b0aa6075d164df9ae4766876cc823394abaebc6d
[karo-tx-linux.git] / mm / shmem.c
1 /*
2  * Resizable virtual memory filesystem for Linux.
3  *
4  * Copyright (C) 2000 Linus Torvalds.
5  *               2000 Transmeta Corp.
6  *               2000-2001 Christoph Rohland
7  *               2000-2001 SAP AG
8  *               2002 Red Hat Inc.
9  * Copyright (C) 2002-2011 Hugh Dickins.
10  * Copyright (C) 2011 Google Inc.
11  * Copyright (C) 2002-2005 VERITAS Software Corporation.
12  * Copyright (C) 2004 Andi Kleen, SuSE Labs
13  *
14  * Extended attribute support for tmpfs:
15  * Copyright (c) 2004, Luke Kenneth Casson Leighton <lkcl@lkcl.net>
16  * Copyright (c) 2004 Red Hat, Inc., James Morris <jmorris@redhat.com>
17  *
18  * tiny-shmem:
19  * Copyright (c) 2004, 2008 Matt Mackall <mpm@selenic.com>
20  *
21  * This file is released under the GPL.
22  */
23
24 #include <linux/fs.h>
25 #include <linux/init.h>
26 #include <linux/vfs.h>
27 #include <linux/mount.h>
28 #include <linux/ramfs.h>
29 #include <linux/pagemap.h>
30 #include <linux/file.h>
31 #include <linux/mm.h>
32 #include <linux/sched/signal.h>
33 #include <linux/export.h>
34 #include <linux/swap.h>
35 #include <linux/uio.h>
36 #include <linux/khugepaged.h>
37
38 #include <asm/tlbflush.h> /* for arch/microblaze update_mmu_cache() */
39
40 static struct vfsmount *shm_mnt;
41
42 #ifdef CONFIG_SHMEM
43 /*
44  * This virtual memory filesystem is heavily based on the ramfs. It
45  * extends ramfs by the ability to use swap and honor resource limits
46  * which makes it a completely usable filesystem.
47  */
48
49 #include <linux/xattr.h>
50 #include <linux/exportfs.h>
51 #include <linux/posix_acl.h>
52 #include <linux/posix_acl_xattr.h>
53 #include <linux/mman.h>
54 #include <linux/string.h>
55 #include <linux/slab.h>
56 #include <linux/backing-dev.h>
57 #include <linux/shmem_fs.h>
58 #include <linux/writeback.h>
59 #include <linux/blkdev.h>
60 #include <linux/pagevec.h>
61 #include <linux/percpu_counter.h>
62 #include <linux/falloc.h>
63 #include <linux/splice.h>
64 #include <linux/security.h>
65 #include <linux/swapops.h>
66 #include <linux/mempolicy.h>
67 #include <linux/namei.h>
68 #include <linux/ctype.h>
69 #include <linux/migrate.h>
70 #include <linux/highmem.h>
71 #include <linux/seq_file.h>
72 #include <linux/magic.h>
73 #include <linux/syscalls.h>
74 #include <linux/fcntl.h>
75 #include <uapi/linux/memfd.h>
76 #include <linux/userfaultfd_k.h>
77 #include <linux/rmap.h>
78 #include <linux/uuid.h>
79
80 #include <linux/uaccess.h>
81 #include <asm/pgtable.h>
82
83 #include "internal.h"
84
85 #define BLOCKS_PER_PAGE  (PAGE_SIZE/512)
86 #define VM_ACCT(size)    (PAGE_ALIGN(size) >> PAGE_SHIFT)
87
88 /* Pretend that each entry is of this size in directory's i_size */
89 #define BOGO_DIRENT_SIZE 20
90
91 /* Symlink up to this size is kmalloc'ed instead of using a swappable page */
92 #define SHORT_SYMLINK_LEN 128
93
94 /*
95  * shmem_fallocate communicates with shmem_fault or shmem_writepage via
96  * inode->i_private (with i_mutex making sure that it has only one user at
97  * a time): we would prefer not to enlarge the shmem inode just for that.
98  */
99 struct shmem_falloc {
100         wait_queue_head_t *waitq; /* faults into hole wait for punch to end */
101         pgoff_t start;          /* start of range currently being fallocated */
102         pgoff_t next;           /* the next page offset to be fallocated */
103         pgoff_t nr_falloced;    /* how many new pages have been fallocated */
104         pgoff_t nr_unswapped;   /* how often writepage refused to swap out */
105 };
106
107 #ifdef CONFIG_TMPFS
108 static unsigned long shmem_default_max_blocks(void)
109 {
110         return totalram_pages / 2;
111 }
112
113 static unsigned long shmem_default_max_inodes(void)
114 {
115         return min(totalram_pages - totalhigh_pages, totalram_pages / 2);
116 }
117 #endif
118
119 static bool shmem_should_replace_page(struct page *page, gfp_t gfp);
120 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
121                                 struct shmem_inode_info *info, pgoff_t index);
122 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
123                 struct page **pagep, enum sgp_type sgp,
124                 gfp_t gfp, struct vm_area_struct *vma,
125                 struct vm_fault *vmf, int *fault_type);
126
127 int shmem_getpage(struct inode *inode, pgoff_t index,
128                 struct page **pagep, enum sgp_type sgp)
129 {
130         return shmem_getpage_gfp(inode, index, pagep, sgp,
131                 mapping_gfp_mask(inode->i_mapping), NULL, NULL, NULL);
132 }
133
134 static inline struct shmem_sb_info *SHMEM_SB(struct super_block *sb)
135 {
136         return sb->s_fs_info;
137 }
138
139 /*
140  * shmem_file_setup pre-accounts the whole fixed size of a VM object,
141  * for shared memory and for shared anonymous (/dev/zero) mappings
142  * (unless MAP_NORESERVE and sysctl_overcommit_memory <= 1),
143  * consistent with the pre-accounting of private mappings ...
144  */
145 static inline int shmem_acct_size(unsigned long flags, loff_t size)
146 {
147         return (flags & VM_NORESERVE) ?
148                 0 : security_vm_enough_memory_mm(current->mm, VM_ACCT(size));
149 }
150
151 static inline void shmem_unacct_size(unsigned long flags, loff_t size)
152 {
153         if (!(flags & VM_NORESERVE))
154                 vm_unacct_memory(VM_ACCT(size));
155 }
156
157 static inline int shmem_reacct_size(unsigned long flags,
158                 loff_t oldsize, loff_t newsize)
159 {
160         if (!(flags & VM_NORESERVE)) {
161                 if (VM_ACCT(newsize) > VM_ACCT(oldsize))
162                         return security_vm_enough_memory_mm(current->mm,
163                                         VM_ACCT(newsize) - VM_ACCT(oldsize));
164                 else if (VM_ACCT(newsize) < VM_ACCT(oldsize))
165                         vm_unacct_memory(VM_ACCT(oldsize) - VM_ACCT(newsize));
166         }
167         return 0;
168 }
169
170 /*
171  * ... whereas tmpfs objects are accounted incrementally as
172  * pages are allocated, in order to allow large sparse files.
173  * shmem_getpage reports shmem_acct_block failure as -ENOSPC not -ENOMEM,
174  * so that a failure on a sparse tmpfs mapping will give SIGBUS not OOM.
175  */
176 static inline int shmem_acct_block(unsigned long flags, long pages)
177 {
178         if (!(flags & VM_NORESERVE))
179                 return 0;
180
181         return security_vm_enough_memory_mm(current->mm,
182                         pages * VM_ACCT(PAGE_SIZE));
183 }
184
185 static inline void shmem_unacct_blocks(unsigned long flags, long pages)
186 {
187         if (flags & VM_NORESERVE)
188                 vm_unacct_memory(pages * VM_ACCT(PAGE_SIZE));
189 }
190
191 static const struct super_operations shmem_ops;
192 static const struct address_space_operations shmem_aops;
193 static const struct file_operations shmem_file_operations;
194 static const struct inode_operations shmem_inode_operations;
195 static const struct inode_operations shmem_dir_inode_operations;
196 static const struct inode_operations shmem_special_inode_operations;
197 static const struct vm_operations_struct shmem_vm_ops;
198 static struct file_system_type shmem_fs_type;
199
200 bool vma_is_shmem(struct vm_area_struct *vma)
201 {
202         return vma->vm_ops == &shmem_vm_ops;
203 }
204
205 static LIST_HEAD(shmem_swaplist);
206 static DEFINE_MUTEX(shmem_swaplist_mutex);
207
208 static int shmem_reserve_inode(struct super_block *sb)
209 {
210         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
211         if (sbinfo->max_inodes) {
212                 spin_lock(&sbinfo->stat_lock);
213                 if (!sbinfo->free_inodes) {
214                         spin_unlock(&sbinfo->stat_lock);
215                         return -ENOSPC;
216                 }
217                 sbinfo->free_inodes--;
218                 spin_unlock(&sbinfo->stat_lock);
219         }
220         return 0;
221 }
222
223 static void shmem_free_inode(struct super_block *sb)
224 {
225         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
226         if (sbinfo->max_inodes) {
227                 spin_lock(&sbinfo->stat_lock);
228                 sbinfo->free_inodes++;
229                 spin_unlock(&sbinfo->stat_lock);
230         }
231 }
232
233 /**
234  * shmem_recalc_inode - recalculate the block usage of an inode
235  * @inode: inode to recalc
236  *
237  * We have to calculate the free blocks since the mm can drop
238  * undirtied hole pages behind our back.
239  *
240  * But normally   info->alloced == inode->i_mapping->nrpages + info->swapped
241  * So mm freed is info->alloced - (inode->i_mapping->nrpages + info->swapped)
242  *
243  * It has to be called with the spinlock held.
244  */
245 static void shmem_recalc_inode(struct inode *inode)
246 {
247         struct shmem_inode_info *info = SHMEM_I(inode);
248         long freed;
249
250         freed = info->alloced - info->swapped - inode->i_mapping->nrpages;
251         if (freed > 0) {
252                 struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
253                 if (sbinfo->max_blocks)
254                         percpu_counter_add(&sbinfo->used_blocks, -freed);
255                 info->alloced -= freed;
256                 inode->i_blocks -= freed * BLOCKS_PER_PAGE;
257                 shmem_unacct_blocks(info->flags, freed);
258         }
259 }
260
261 bool shmem_charge(struct inode *inode, long pages)
262 {
263         struct shmem_inode_info *info = SHMEM_I(inode);
264         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
265         unsigned long flags;
266
267         if (shmem_acct_block(info->flags, pages))
268                 return false;
269         spin_lock_irqsave(&info->lock, flags);
270         info->alloced += pages;
271         inode->i_blocks += pages * BLOCKS_PER_PAGE;
272         shmem_recalc_inode(inode);
273         spin_unlock_irqrestore(&info->lock, flags);
274         inode->i_mapping->nrpages += pages;
275
276         if (!sbinfo->max_blocks)
277                 return true;
278         if (percpu_counter_compare(&sbinfo->used_blocks,
279                                 sbinfo->max_blocks - pages) > 0) {
280                 inode->i_mapping->nrpages -= pages;
281                 spin_lock_irqsave(&info->lock, flags);
282                 info->alloced -= pages;
283                 shmem_recalc_inode(inode);
284                 spin_unlock_irqrestore(&info->lock, flags);
285                 shmem_unacct_blocks(info->flags, pages);
286                 return false;
287         }
288         percpu_counter_add(&sbinfo->used_blocks, pages);
289         return true;
290 }
291
292 void shmem_uncharge(struct inode *inode, long pages)
293 {
294         struct shmem_inode_info *info = SHMEM_I(inode);
295         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
296         unsigned long flags;
297
298         spin_lock_irqsave(&info->lock, flags);
299         info->alloced -= pages;
300         inode->i_blocks -= pages * BLOCKS_PER_PAGE;
301         shmem_recalc_inode(inode);
302         spin_unlock_irqrestore(&info->lock, flags);
303
304         if (sbinfo->max_blocks)
305                 percpu_counter_sub(&sbinfo->used_blocks, pages);
306         shmem_unacct_blocks(info->flags, pages);
307 }
308
309 /*
310  * Replace item expected in radix tree by a new item, while holding tree lock.
311  */
312 static int shmem_radix_tree_replace(struct address_space *mapping,
313                         pgoff_t index, void *expected, void *replacement)
314 {
315         struct radix_tree_node *node;
316         void **pslot;
317         void *item;
318
319         VM_BUG_ON(!expected);
320         VM_BUG_ON(!replacement);
321         item = __radix_tree_lookup(&mapping->page_tree, index, &node, &pslot);
322         if (!item)
323                 return -ENOENT;
324         if (item != expected)
325                 return -ENOENT;
326         __radix_tree_replace(&mapping->page_tree, node, pslot,
327                              replacement, NULL, NULL);
328         return 0;
329 }
330
331 /*
332  * Sometimes, before we decide whether to proceed or to fail, we must check
333  * that an entry was not already brought back from swap by a racing thread.
334  *
335  * Checking page is not enough: by the time a SwapCache page is locked, it
336  * might be reused, and again be SwapCache, using the same swap as before.
337  */
338 static bool shmem_confirm_swap(struct address_space *mapping,
339                                pgoff_t index, swp_entry_t swap)
340 {
341         void *item;
342
343         rcu_read_lock();
344         item = radix_tree_lookup(&mapping->page_tree, index);
345         rcu_read_unlock();
346         return item == swp_to_radix_entry(swap);
347 }
348
349 /*
350  * Definitions for "huge tmpfs": tmpfs mounted with the huge= option
351  *
352  * SHMEM_HUGE_NEVER:
353  *      disables huge pages for the mount;
354  * SHMEM_HUGE_ALWAYS:
355  *      enables huge pages for the mount;
356  * SHMEM_HUGE_WITHIN_SIZE:
357  *      only allocate huge pages if the page will be fully within i_size,
358  *      also respect fadvise()/madvise() hints;
359  * SHMEM_HUGE_ADVISE:
360  *      only allocate huge pages if requested with fadvise()/madvise();
361  */
362
363 #define SHMEM_HUGE_NEVER        0
364 #define SHMEM_HUGE_ALWAYS       1
365 #define SHMEM_HUGE_WITHIN_SIZE  2
366 #define SHMEM_HUGE_ADVISE       3
367
368 /*
369  * Special values.
370  * Only can be set via /sys/kernel/mm/transparent_hugepage/shmem_enabled:
371  *
372  * SHMEM_HUGE_DENY:
373  *      disables huge on shm_mnt and all mounts, for emergency use;
374  * SHMEM_HUGE_FORCE:
375  *      enables huge on shm_mnt and all mounts, w/o needing option, for testing;
376  *
377  */
378 #define SHMEM_HUGE_DENY         (-1)
379 #define SHMEM_HUGE_FORCE        (-2)
380
381 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
382 /* ifdef here to avoid bloating shmem.o when not necessary */
383
384 int shmem_huge __read_mostly;
385
386 #if defined(CONFIG_SYSFS) || defined(CONFIG_TMPFS)
387 static int shmem_parse_huge(const char *str)
388 {
389         if (!strcmp(str, "never"))
390                 return SHMEM_HUGE_NEVER;
391         if (!strcmp(str, "always"))
392                 return SHMEM_HUGE_ALWAYS;
393         if (!strcmp(str, "within_size"))
394                 return SHMEM_HUGE_WITHIN_SIZE;
395         if (!strcmp(str, "advise"))
396                 return SHMEM_HUGE_ADVISE;
397         if (!strcmp(str, "deny"))
398                 return SHMEM_HUGE_DENY;
399         if (!strcmp(str, "force"))
400                 return SHMEM_HUGE_FORCE;
401         return -EINVAL;
402 }
403
404 static const char *shmem_format_huge(int huge)
405 {
406         switch (huge) {
407         case SHMEM_HUGE_NEVER:
408                 return "never";
409         case SHMEM_HUGE_ALWAYS:
410                 return "always";
411         case SHMEM_HUGE_WITHIN_SIZE:
412                 return "within_size";
413         case SHMEM_HUGE_ADVISE:
414                 return "advise";
415         case SHMEM_HUGE_DENY:
416                 return "deny";
417         case SHMEM_HUGE_FORCE:
418                 return "force";
419         default:
420                 VM_BUG_ON(1);
421                 return "bad_val";
422         }
423 }
424 #endif
425
426 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
427                 struct shrink_control *sc, unsigned long nr_to_split)
428 {
429         LIST_HEAD(list), *pos, *next;
430         LIST_HEAD(to_remove);
431         struct inode *inode;
432         struct shmem_inode_info *info;
433         struct page *page;
434         unsigned long batch = sc ? sc->nr_to_scan : 128;
435         int removed = 0, split = 0;
436
437         if (list_empty(&sbinfo->shrinklist))
438                 return SHRINK_STOP;
439
440         spin_lock(&sbinfo->shrinklist_lock);
441         list_for_each_safe(pos, next, &sbinfo->shrinklist) {
442                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
443
444                 /* pin the inode */
445                 inode = igrab(&info->vfs_inode);
446
447                 /* inode is about to be evicted */
448                 if (!inode) {
449                         list_del_init(&info->shrinklist);
450                         removed++;
451                         goto next;
452                 }
453
454                 /* Check if there's anything to gain */
455                 if (round_up(inode->i_size, PAGE_SIZE) ==
456                                 round_up(inode->i_size, HPAGE_PMD_SIZE)) {
457                         list_move(&info->shrinklist, &to_remove);
458                         removed++;
459                         goto next;
460                 }
461
462                 list_move(&info->shrinklist, &list);
463 next:
464                 if (!--batch)
465                         break;
466         }
467         spin_unlock(&sbinfo->shrinklist_lock);
468
469         list_for_each_safe(pos, next, &to_remove) {
470                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
471                 inode = &info->vfs_inode;
472                 list_del_init(&info->shrinklist);
473                 iput(inode);
474         }
475
476         list_for_each_safe(pos, next, &list) {
477                 int ret;
478
479                 info = list_entry(pos, struct shmem_inode_info, shrinklist);
480                 inode = &info->vfs_inode;
481
482                 if (nr_to_split && split >= nr_to_split) {
483                         iput(inode);
484                         continue;
485                 }
486
487                 page = find_lock_page(inode->i_mapping,
488                                 (inode->i_size & HPAGE_PMD_MASK) >> PAGE_SHIFT);
489                 if (!page)
490                         goto drop;
491
492                 if (!PageTransHuge(page)) {
493                         unlock_page(page);
494                         put_page(page);
495                         goto drop;
496                 }
497
498                 ret = split_huge_page(page);
499                 unlock_page(page);
500                 put_page(page);
501
502                 if (ret) {
503                         /* split failed: leave it on the list */
504                         iput(inode);
505                         continue;
506                 }
507
508                 split++;
509 drop:
510                 list_del_init(&info->shrinklist);
511                 removed++;
512                 iput(inode);
513         }
514
515         spin_lock(&sbinfo->shrinklist_lock);
516         list_splice_tail(&list, &sbinfo->shrinklist);
517         sbinfo->shrinklist_len -= removed;
518         spin_unlock(&sbinfo->shrinklist_lock);
519
520         return split;
521 }
522
523 static long shmem_unused_huge_scan(struct super_block *sb,
524                 struct shrink_control *sc)
525 {
526         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
527
528         if (!READ_ONCE(sbinfo->shrinklist_len))
529                 return SHRINK_STOP;
530
531         return shmem_unused_huge_shrink(sbinfo, sc, 0);
532 }
533
534 static long shmem_unused_huge_count(struct super_block *sb,
535                 struct shrink_control *sc)
536 {
537         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
538         return READ_ONCE(sbinfo->shrinklist_len);
539 }
540 #else /* !CONFIG_TRANSPARENT_HUGE_PAGECACHE */
541
542 #define shmem_huge SHMEM_HUGE_DENY
543
544 static unsigned long shmem_unused_huge_shrink(struct shmem_sb_info *sbinfo,
545                 struct shrink_control *sc, unsigned long nr_to_split)
546 {
547         return 0;
548 }
549 #endif /* CONFIG_TRANSPARENT_HUGE_PAGECACHE */
550
551 /*
552  * Like add_to_page_cache_locked, but error if expected item has gone.
553  */
554 static int shmem_add_to_page_cache(struct page *page,
555                                    struct address_space *mapping,
556                                    pgoff_t index, void *expected)
557 {
558         int error, nr = hpage_nr_pages(page);
559
560         VM_BUG_ON_PAGE(PageTail(page), page);
561         VM_BUG_ON_PAGE(index != round_down(index, nr), page);
562         VM_BUG_ON_PAGE(!PageLocked(page), page);
563         VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
564         VM_BUG_ON(expected && PageTransHuge(page));
565
566         page_ref_add(page, nr);
567         page->mapping = mapping;
568         page->index = index;
569
570         spin_lock_irq(&mapping->tree_lock);
571         if (PageTransHuge(page)) {
572                 void __rcu **results;
573                 pgoff_t idx;
574                 int i;
575
576                 error = 0;
577                 if (radix_tree_gang_lookup_slot(&mapping->page_tree,
578                                         &results, &idx, index, 1) &&
579                                 idx < index + HPAGE_PMD_NR) {
580                         error = -EEXIST;
581                 }
582
583                 if (!error) {
584                         for (i = 0; i < HPAGE_PMD_NR; i++) {
585                                 error = radix_tree_insert(&mapping->page_tree,
586                                                 index + i, page + i);
587                                 VM_BUG_ON(error);
588                         }
589                         count_vm_event(THP_FILE_ALLOC);
590                 }
591         } else if (!expected) {
592                 error = radix_tree_insert(&mapping->page_tree, index, page);
593         } else {
594                 error = shmem_radix_tree_replace(mapping, index, expected,
595                                                                  page);
596         }
597
598         if (!error) {
599                 mapping->nrpages += nr;
600                 if (PageTransHuge(page))
601                         __inc_node_page_state(page, NR_SHMEM_THPS);
602                 __mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
603                 __mod_node_page_state(page_pgdat(page), NR_SHMEM, nr);
604                 spin_unlock_irq(&mapping->tree_lock);
605         } else {
606                 page->mapping = NULL;
607                 spin_unlock_irq(&mapping->tree_lock);
608                 page_ref_sub(page, nr);
609         }
610         return error;
611 }
612
613 /*
614  * Like delete_from_page_cache, but substitutes swap for page.
615  */
616 static void shmem_delete_from_page_cache(struct page *page, void *radswap)
617 {
618         struct address_space *mapping = page->mapping;
619         int error;
620
621         VM_BUG_ON_PAGE(PageCompound(page), page);
622
623         spin_lock_irq(&mapping->tree_lock);
624         error = shmem_radix_tree_replace(mapping, page->index, page, radswap);
625         page->mapping = NULL;
626         mapping->nrpages--;
627         __dec_node_page_state(page, NR_FILE_PAGES);
628         __dec_node_page_state(page, NR_SHMEM);
629         spin_unlock_irq(&mapping->tree_lock);
630         put_page(page);
631         BUG_ON(error);
632 }
633
634 /*
635  * Remove swap entry from radix tree, free the swap and its page cache.
636  */
637 static int shmem_free_swap(struct address_space *mapping,
638                            pgoff_t index, void *radswap)
639 {
640         void *old;
641
642         spin_lock_irq(&mapping->tree_lock);
643         old = radix_tree_delete_item(&mapping->page_tree, index, radswap);
644         spin_unlock_irq(&mapping->tree_lock);
645         if (old != radswap)
646                 return -ENOENT;
647         free_swap_and_cache(radix_to_swp_entry(radswap));
648         return 0;
649 }
650
651 /*
652  * Determine (in bytes) how many of the shmem object's pages mapped by the
653  * given offsets are swapped out.
654  *
655  * This is safe to call without i_mutex or mapping->tree_lock thanks to RCU,
656  * as long as the inode doesn't go away and racy results are not a problem.
657  */
658 unsigned long shmem_partial_swap_usage(struct address_space *mapping,
659                                                 pgoff_t start, pgoff_t end)
660 {
661         struct radix_tree_iter iter;
662         void **slot;
663         struct page *page;
664         unsigned long swapped = 0;
665
666         rcu_read_lock();
667
668         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
669                 if (iter.index >= end)
670                         break;
671
672                 page = radix_tree_deref_slot(slot);
673
674                 if (radix_tree_deref_retry(page)) {
675                         slot = radix_tree_iter_retry(&iter);
676                         continue;
677                 }
678
679                 if (radix_tree_exceptional_entry(page))
680                         swapped++;
681
682                 if (need_resched()) {
683                         slot = radix_tree_iter_resume(slot, &iter);
684                         cond_resched_rcu();
685                 }
686         }
687
688         rcu_read_unlock();
689
690         return swapped << PAGE_SHIFT;
691 }
692
693 /*
694  * Determine (in bytes) how many of the shmem object's pages mapped by the
695  * given vma is swapped out.
696  *
697  * This is safe to call without i_mutex or mapping->tree_lock thanks to RCU,
698  * as long as the inode doesn't go away and racy results are not a problem.
699  */
700 unsigned long shmem_swap_usage(struct vm_area_struct *vma)
701 {
702         struct inode *inode = file_inode(vma->vm_file);
703         struct shmem_inode_info *info = SHMEM_I(inode);
704         struct address_space *mapping = inode->i_mapping;
705         unsigned long swapped;
706
707         /* Be careful as we don't hold info->lock */
708         swapped = READ_ONCE(info->swapped);
709
710         /*
711          * The easier cases are when the shmem object has nothing in swap, or
712          * the vma maps it whole. Then we can simply use the stats that we
713          * already track.
714          */
715         if (!swapped)
716                 return 0;
717
718         if (!vma->vm_pgoff && vma->vm_end - vma->vm_start >= inode->i_size)
719                 return swapped << PAGE_SHIFT;
720
721         /* Here comes the more involved part */
722         return shmem_partial_swap_usage(mapping,
723                         linear_page_index(vma, vma->vm_start),
724                         linear_page_index(vma, vma->vm_end));
725 }
726
727 /*
728  * SysV IPC SHM_UNLOCK restore Unevictable pages to their evictable lists.
729  */
730 void shmem_unlock_mapping(struct address_space *mapping)
731 {
732         struct pagevec pvec;
733         pgoff_t indices[PAGEVEC_SIZE];
734         pgoff_t index = 0;
735
736         pagevec_init(&pvec, 0);
737         /*
738          * Minor point, but we might as well stop if someone else SHM_LOCKs it.
739          */
740         while (!mapping_unevictable(mapping)) {
741                 /*
742                  * Avoid pagevec_lookup(): find_get_pages() returns 0 as if it
743                  * has finished, if it hits a row of PAGEVEC_SIZE swap entries.
744                  */
745                 pvec.nr = find_get_entries(mapping, index,
746                                            PAGEVEC_SIZE, pvec.pages, indices);
747                 if (!pvec.nr)
748                         break;
749                 index = indices[pvec.nr - 1] + 1;
750                 pagevec_remove_exceptionals(&pvec);
751                 check_move_unevictable_pages(pvec.pages, pvec.nr);
752                 pagevec_release(&pvec);
753                 cond_resched();
754         }
755 }
756
757 /*
758  * Remove range of pages and swap entries from radix tree, and free them.
759  * If !unfalloc, truncate or punch hole; if unfalloc, undo failed fallocate.
760  */
761 static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
762                                                                  bool unfalloc)
763 {
764         struct address_space *mapping = inode->i_mapping;
765         struct shmem_inode_info *info = SHMEM_I(inode);
766         pgoff_t start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
767         pgoff_t end = (lend + 1) >> PAGE_SHIFT;
768         unsigned int partial_start = lstart & (PAGE_SIZE - 1);
769         unsigned int partial_end = (lend + 1) & (PAGE_SIZE - 1);
770         struct pagevec pvec;
771         pgoff_t indices[PAGEVEC_SIZE];
772         long nr_swaps_freed = 0;
773         pgoff_t index;
774         int i;
775
776         if (lend == -1)
777                 end = -1;       /* unsigned, so actually very big */
778
779         pagevec_init(&pvec, 0);
780         index = start;
781         while (index < end) {
782                 pvec.nr = find_get_entries(mapping, index,
783                         min(end - index, (pgoff_t)PAGEVEC_SIZE),
784                         pvec.pages, indices);
785                 if (!pvec.nr)
786                         break;
787                 for (i = 0; i < pagevec_count(&pvec); i++) {
788                         struct page *page = pvec.pages[i];
789
790                         index = indices[i];
791                         if (index >= end)
792                                 break;
793
794                         if (radix_tree_exceptional_entry(page)) {
795                                 if (unfalloc)
796                                         continue;
797                                 nr_swaps_freed += !shmem_free_swap(mapping,
798                                                                 index, page);
799                                 continue;
800                         }
801
802                         VM_BUG_ON_PAGE(page_to_pgoff(page) != index, page);
803
804                         if (!trylock_page(page))
805                                 continue;
806
807                         if (PageTransTail(page)) {
808                                 /* Middle of THP: zero out the page */
809                                 clear_highpage(page);
810                                 unlock_page(page);
811                                 continue;
812                         } else if (PageTransHuge(page)) {
813                                 if (index == round_down(end, HPAGE_PMD_NR)) {
814                                         /*
815                                          * Range ends in the middle of THP:
816                                          * zero out the page
817                                          */
818                                         clear_highpage(page);
819                                         unlock_page(page);
820                                         continue;
821                                 }
822                                 index += HPAGE_PMD_NR - 1;
823                                 i += HPAGE_PMD_NR - 1;
824                         }
825
826                         if (!unfalloc || !PageUptodate(page)) {
827                                 VM_BUG_ON_PAGE(PageTail(page), page);
828                                 if (page_mapping(page) == mapping) {
829                                         VM_BUG_ON_PAGE(PageWriteback(page), page);
830                                         truncate_inode_page(mapping, page);
831                                 }
832                         }
833                         unlock_page(page);
834                 }
835                 pagevec_remove_exceptionals(&pvec);
836                 pagevec_release(&pvec);
837                 cond_resched();
838                 index++;
839         }
840
841         if (partial_start) {
842                 struct page *page = NULL;
843                 shmem_getpage(inode, start - 1, &page, SGP_READ);
844                 if (page) {
845                         unsigned int top = PAGE_SIZE;
846                         if (start > end) {
847                                 top = partial_end;
848                                 partial_end = 0;
849                         }
850                         zero_user_segment(page, partial_start, top);
851                         set_page_dirty(page);
852                         unlock_page(page);
853                         put_page(page);
854                 }
855         }
856         if (partial_end) {
857                 struct page *page = NULL;
858                 shmem_getpage(inode, end, &page, SGP_READ);
859                 if (page) {
860                         zero_user_segment(page, 0, partial_end);
861                         set_page_dirty(page);
862                         unlock_page(page);
863                         put_page(page);
864                 }
865         }
866         if (start >= end)
867                 return;
868
869         index = start;
870         while (index < end) {
871                 cond_resched();
872
873                 pvec.nr = find_get_entries(mapping, index,
874                                 min(end - index, (pgoff_t)PAGEVEC_SIZE),
875                                 pvec.pages, indices);
876                 if (!pvec.nr) {
877                         /* If all gone or hole-punch or unfalloc, we're done */
878                         if (index == start || end != -1)
879                                 break;
880                         /* But if truncating, restart to make sure all gone */
881                         index = start;
882                         continue;
883                 }
884                 for (i = 0; i < pagevec_count(&pvec); i++) {
885                         struct page *page = pvec.pages[i];
886
887                         index = indices[i];
888                         if (index >= end)
889                                 break;
890
891                         if (radix_tree_exceptional_entry(page)) {
892                                 if (unfalloc)
893                                         continue;
894                                 if (shmem_free_swap(mapping, index, page)) {
895                                         /* Swap was replaced by page: retry */
896                                         index--;
897                                         break;
898                                 }
899                                 nr_swaps_freed++;
900                                 continue;
901                         }
902
903                         lock_page(page);
904
905                         if (PageTransTail(page)) {
906                                 /* Middle of THP: zero out the page */
907                                 clear_highpage(page);
908                                 unlock_page(page);
909                                 /*
910                                  * Partial thp truncate due 'start' in middle
911                                  * of THP: don't need to look on these pages
912                                  * again on !pvec.nr restart.
913                                  */
914                                 if (index != round_down(end, HPAGE_PMD_NR))
915                                         start++;
916                                 continue;
917                         } else if (PageTransHuge(page)) {
918                                 if (index == round_down(end, HPAGE_PMD_NR)) {
919                                         /*
920                                          * Range ends in the middle of THP:
921                                          * zero out the page
922                                          */
923                                         clear_highpage(page);
924                                         unlock_page(page);
925                                         continue;
926                                 }
927                                 index += HPAGE_PMD_NR - 1;
928                                 i += HPAGE_PMD_NR - 1;
929                         }
930
931                         if (!unfalloc || !PageUptodate(page)) {
932                                 VM_BUG_ON_PAGE(PageTail(page), page);
933                                 if (page_mapping(page) == mapping) {
934                                         VM_BUG_ON_PAGE(PageWriteback(page), page);
935                                         truncate_inode_page(mapping, page);
936                                 } else {
937                                         /* Page was replaced by swap: retry */
938                                         unlock_page(page);
939                                         index--;
940                                         break;
941                                 }
942                         }
943                         unlock_page(page);
944                 }
945                 pagevec_remove_exceptionals(&pvec);
946                 pagevec_release(&pvec);
947                 index++;
948         }
949
950         spin_lock_irq(&info->lock);
951         info->swapped -= nr_swaps_freed;
952         shmem_recalc_inode(inode);
953         spin_unlock_irq(&info->lock);
954 }
955
956 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
957 {
958         shmem_undo_range(inode, lstart, lend, false);
959         inode->i_ctime = inode->i_mtime = current_time(inode);
960 }
961 EXPORT_SYMBOL_GPL(shmem_truncate_range);
962
963 static int shmem_getattr(const struct path *path, struct kstat *stat,
964                          u32 request_mask, unsigned int query_flags)
965 {
966         struct inode *inode = path->dentry->d_inode;
967         struct shmem_inode_info *info = SHMEM_I(inode);
968
969         if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
970                 spin_lock_irq(&info->lock);
971                 shmem_recalc_inode(inode);
972                 spin_unlock_irq(&info->lock);
973         }
974         generic_fillattr(inode, stat);
975         return 0;
976 }
977
978 static int shmem_setattr(struct dentry *dentry, struct iattr *attr)
979 {
980         struct inode *inode = d_inode(dentry);
981         struct shmem_inode_info *info = SHMEM_I(inode);
982         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
983         int error;
984
985         error = setattr_prepare(dentry, attr);
986         if (error)
987                 return error;
988
989         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
990                 loff_t oldsize = inode->i_size;
991                 loff_t newsize = attr->ia_size;
992
993                 /* protected by i_mutex */
994                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
995                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
996                         return -EPERM;
997
998                 if (newsize != oldsize) {
999                         error = shmem_reacct_size(SHMEM_I(inode)->flags,
1000                                         oldsize, newsize);
1001                         if (error)
1002                                 return error;
1003                         i_size_write(inode, newsize);
1004                         inode->i_ctime = inode->i_mtime = current_time(inode);
1005                 }
1006                 if (newsize <= oldsize) {
1007                         loff_t holebegin = round_up(newsize, PAGE_SIZE);
1008                         if (oldsize > holebegin)
1009                                 unmap_mapping_range(inode->i_mapping,
1010                                                         holebegin, 0, 1);
1011                         if (info->alloced)
1012                                 shmem_truncate_range(inode,
1013                                                         newsize, (loff_t)-1);
1014                         /* unmap again to remove racily COWed private pages */
1015                         if (oldsize > holebegin)
1016                                 unmap_mapping_range(inode->i_mapping,
1017                                                         holebegin, 0, 1);
1018
1019                         /*
1020                          * Part of the huge page can be beyond i_size: subject
1021                          * to shrink under memory pressure.
1022                          */
1023                         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
1024                                 spin_lock(&sbinfo->shrinklist_lock);
1025                                 if (list_empty(&info->shrinklist)) {
1026                                         list_add_tail(&info->shrinklist,
1027                                                         &sbinfo->shrinklist);
1028                                         sbinfo->shrinklist_len++;
1029                                 }
1030                                 spin_unlock(&sbinfo->shrinklist_lock);
1031                         }
1032                 }
1033         }
1034
1035         setattr_copy(inode, attr);
1036         if (attr->ia_valid & ATTR_MODE)
1037                 error = posix_acl_chmod(inode, inode->i_mode);
1038         return error;
1039 }
1040
1041 static void shmem_evict_inode(struct inode *inode)
1042 {
1043         struct shmem_inode_info *info = SHMEM_I(inode);
1044         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
1045
1046         if (inode->i_mapping->a_ops == &shmem_aops) {
1047                 shmem_unacct_size(info->flags, inode->i_size);
1048                 inode->i_size = 0;
1049                 shmem_truncate_range(inode, 0, (loff_t)-1);
1050                 if (!list_empty(&info->shrinklist)) {
1051                         spin_lock(&sbinfo->shrinklist_lock);
1052                         if (!list_empty(&info->shrinklist)) {
1053                                 list_del_init(&info->shrinklist);
1054                                 sbinfo->shrinklist_len--;
1055                         }
1056                         spin_unlock(&sbinfo->shrinklist_lock);
1057                 }
1058                 if (!list_empty(&info->swaplist)) {
1059                         mutex_lock(&shmem_swaplist_mutex);
1060                         list_del_init(&info->swaplist);
1061                         mutex_unlock(&shmem_swaplist_mutex);
1062                 }
1063         }
1064
1065         simple_xattrs_free(&info->xattrs);
1066         WARN_ON(inode->i_blocks);
1067         shmem_free_inode(inode->i_sb);
1068         clear_inode(inode);
1069 }
1070
1071 static unsigned long find_swap_entry(struct radix_tree_root *root, void *item)
1072 {
1073         struct radix_tree_iter iter;
1074         void **slot;
1075         unsigned long found = -1;
1076         unsigned int checked = 0;
1077
1078         rcu_read_lock();
1079         radix_tree_for_each_slot(slot, root, &iter, 0) {
1080                 if (*slot == item) {
1081                         found = iter.index;
1082                         break;
1083                 }
1084                 checked++;
1085                 if ((checked % 4096) != 0)
1086                         continue;
1087                 slot = radix_tree_iter_resume(slot, &iter);
1088                 cond_resched_rcu();
1089         }
1090
1091         rcu_read_unlock();
1092         return found;
1093 }
1094
1095 /*
1096  * If swap found in inode, free it and move page from swapcache to filecache.
1097  */
1098 static int shmem_unuse_inode(struct shmem_inode_info *info,
1099                              swp_entry_t swap, struct page **pagep)
1100 {
1101         struct address_space *mapping = info->vfs_inode.i_mapping;
1102         void *radswap;
1103         pgoff_t index;
1104         gfp_t gfp;
1105         int error = 0;
1106
1107         radswap = swp_to_radix_entry(swap);
1108         index = find_swap_entry(&mapping->page_tree, radswap);
1109         if (index == -1)
1110                 return -EAGAIN; /* tell shmem_unuse we found nothing */
1111
1112         /*
1113          * Move _head_ to start search for next from here.
1114          * But be careful: shmem_evict_inode checks list_empty without taking
1115          * mutex, and there's an instant in list_move_tail when info->swaplist
1116          * would appear empty, if it were the only one on shmem_swaplist.
1117          */
1118         if (shmem_swaplist.next != &info->swaplist)
1119                 list_move_tail(&shmem_swaplist, &info->swaplist);
1120
1121         gfp = mapping_gfp_mask(mapping);
1122         if (shmem_should_replace_page(*pagep, gfp)) {
1123                 mutex_unlock(&shmem_swaplist_mutex);
1124                 error = shmem_replace_page(pagep, gfp, info, index);
1125                 mutex_lock(&shmem_swaplist_mutex);
1126                 /*
1127                  * We needed to drop mutex to make that restrictive page
1128                  * allocation, but the inode might have been freed while we
1129                  * dropped it: although a racing shmem_evict_inode() cannot
1130                  * complete without emptying the radix_tree, our page lock
1131                  * on this swapcache page is not enough to prevent that -
1132                  * free_swap_and_cache() of our swap entry will only
1133                  * trylock_page(), removing swap from radix_tree whatever.
1134                  *
1135                  * We must not proceed to shmem_add_to_page_cache() if the
1136                  * inode has been freed, but of course we cannot rely on
1137                  * inode or mapping or info to check that.  However, we can
1138                  * safely check if our swap entry is still in use (and here
1139                  * it can't have got reused for another page): if it's still
1140                  * in use, then the inode cannot have been freed yet, and we
1141                  * can safely proceed (if it's no longer in use, that tells
1142                  * nothing about the inode, but we don't need to unuse swap).
1143                  */
1144                 if (!page_swapcount(*pagep))
1145                         error = -ENOENT;
1146         }
1147
1148         /*
1149          * We rely on shmem_swaplist_mutex, not only to protect the swaplist,
1150          * but also to hold up shmem_evict_inode(): so inode cannot be freed
1151          * beneath us (pagelock doesn't help until the page is in pagecache).
1152          */
1153         if (!error)
1154                 error = shmem_add_to_page_cache(*pagep, mapping, index,
1155                                                 radswap);
1156         if (error != -ENOMEM) {
1157                 /*
1158                  * Truncation and eviction use free_swap_and_cache(), which
1159                  * only does trylock page: if we raced, best clean up here.
1160                  */
1161                 delete_from_swap_cache(*pagep);
1162                 set_page_dirty(*pagep);
1163                 if (!error) {
1164                         spin_lock_irq(&info->lock);
1165                         info->swapped--;
1166                         spin_unlock_irq(&info->lock);
1167                         swap_free(swap);
1168                 }
1169         }
1170         return error;
1171 }
1172
1173 /*
1174  * Search through swapped inodes to find and replace swap by page.
1175  */
1176 int shmem_unuse(swp_entry_t swap, struct page *page)
1177 {
1178         struct list_head *this, *next;
1179         struct shmem_inode_info *info;
1180         struct mem_cgroup *memcg;
1181         int error = 0;
1182
1183         /*
1184          * There's a faint possibility that swap page was replaced before
1185          * caller locked it: caller will come back later with the right page.
1186          */
1187         if (unlikely(!PageSwapCache(page) || page_private(page) != swap.val))
1188                 goto out;
1189
1190         /*
1191          * Charge page using GFP_KERNEL while we can wait, before taking
1192          * the shmem_swaplist_mutex which might hold up shmem_writepage().
1193          * Charged back to the user (not to caller) when swap account is used.
1194          */
1195         error = mem_cgroup_try_charge(page, current->mm, GFP_KERNEL, &memcg,
1196                         false);
1197         if (error)
1198                 goto out;
1199         /* No radix_tree_preload: swap entry keeps a place for page in tree */
1200         error = -EAGAIN;
1201
1202         mutex_lock(&shmem_swaplist_mutex);
1203         list_for_each_safe(this, next, &shmem_swaplist) {
1204                 info = list_entry(this, struct shmem_inode_info, swaplist);
1205                 if (info->swapped)
1206                         error = shmem_unuse_inode(info, swap, &page);
1207                 else
1208                         list_del_init(&info->swaplist);
1209                 cond_resched();
1210                 if (error != -EAGAIN)
1211                         break;
1212                 /* found nothing in this: move on to search the next */
1213         }
1214         mutex_unlock(&shmem_swaplist_mutex);
1215
1216         if (error) {
1217                 if (error != -ENOMEM)
1218                         error = 0;
1219                 mem_cgroup_cancel_charge(page, memcg, false);
1220         } else
1221                 mem_cgroup_commit_charge(page, memcg, true, false);
1222 out:
1223         unlock_page(page);
1224         put_page(page);
1225         return error;
1226 }
1227
1228 /*
1229  * Move the page from the page cache to the swap cache.
1230  */
1231 static int shmem_writepage(struct page *page, struct writeback_control *wbc)
1232 {
1233         struct shmem_inode_info *info;
1234         struct address_space *mapping;
1235         struct inode *inode;
1236         swp_entry_t swap;
1237         pgoff_t index;
1238
1239         VM_BUG_ON_PAGE(PageCompound(page), page);
1240         BUG_ON(!PageLocked(page));
1241         mapping = page->mapping;
1242         index = page->index;
1243         inode = mapping->host;
1244         info = SHMEM_I(inode);
1245         if (info->flags & VM_LOCKED)
1246                 goto redirty;
1247         if (!total_swap_pages)
1248                 goto redirty;
1249
1250         /*
1251          * Our capabilities prevent regular writeback or sync from ever calling
1252          * shmem_writepage; but a stacking filesystem might use ->writepage of
1253          * its underlying filesystem, in which case tmpfs should write out to
1254          * swap only in response to memory pressure, and not for the writeback
1255          * threads or sync.
1256          */
1257         if (!wbc->for_reclaim) {
1258                 WARN_ON_ONCE(1);        /* Still happens? Tell us about it! */
1259                 goto redirty;
1260         }
1261
1262         /*
1263          * This is somewhat ridiculous, but without plumbing a SWAP_MAP_FALLOC
1264          * value into swapfile.c, the only way we can correctly account for a
1265          * fallocated page arriving here is now to initialize it and write it.
1266          *
1267          * That's okay for a page already fallocated earlier, but if we have
1268          * not yet completed the fallocation, then (a) we want to keep track
1269          * of this page in case we have to undo it, and (b) it may not be a
1270          * good idea to continue anyway, once we're pushing into swap.  So
1271          * reactivate the page, and let shmem_fallocate() quit when too many.
1272          */
1273         if (!PageUptodate(page)) {
1274                 if (inode->i_private) {
1275                         struct shmem_falloc *shmem_falloc;
1276                         spin_lock(&inode->i_lock);
1277                         shmem_falloc = inode->i_private;
1278                         if (shmem_falloc &&
1279                             !shmem_falloc->waitq &&
1280                             index >= shmem_falloc->start &&
1281                             index < shmem_falloc->next)
1282                                 shmem_falloc->nr_unswapped++;
1283                         else
1284                                 shmem_falloc = NULL;
1285                         spin_unlock(&inode->i_lock);
1286                         if (shmem_falloc)
1287                                 goto redirty;
1288                 }
1289                 clear_highpage(page);
1290                 flush_dcache_page(page);
1291                 SetPageUptodate(page);
1292         }
1293
1294         swap = get_swap_page(page);
1295         if (!swap.val)
1296                 goto redirty;
1297
1298         if (mem_cgroup_try_charge_swap(page, swap))
1299                 goto free_swap;
1300
1301         /*
1302          * Add inode to shmem_unuse()'s list of swapped-out inodes,
1303          * if it's not already there.  Do it now before the page is
1304          * moved to swap cache, when its pagelock no longer protects
1305          * the inode from eviction.  But don't unlock the mutex until
1306          * we've incremented swapped, because shmem_unuse_inode() will
1307          * prune a !swapped inode from the swaplist under this mutex.
1308          */
1309         mutex_lock(&shmem_swaplist_mutex);
1310         if (list_empty(&info->swaplist))
1311                 list_add_tail(&info->swaplist, &shmem_swaplist);
1312
1313         if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
1314                 spin_lock_irq(&info->lock);
1315                 shmem_recalc_inode(inode);
1316                 info->swapped++;
1317                 spin_unlock_irq(&info->lock);
1318
1319                 swap_shmem_alloc(swap);
1320                 shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
1321
1322                 mutex_unlock(&shmem_swaplist_mutex);
1323                 BUG_ON(page_mapped(page));
1324                 swap_writepage(page, wbc);
1325                 return 0;
1326         }
1327
1328         mutex_unlock(&shmem_swaplist_mutex);
1329 free_swap:
1330         put_swap_page(page, swap);
1331 redirty:
1332         set_page_dirty(page);
1333         if (wbc->for_reclaim)
1334                 return AOP_WRITEPAGE_ACTIVATE;  /* Return with page locked */
1335         unlock_page(page);
1336         return 0;
1337 }
1338
1339 #if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)
1340 static void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1341 {
1342         char buffer[64];
1343
1344         if (!mpol || mpol->mode == MPOL_DEFAULT)
1345                 return;         /* show nothing */
1346
1347         mpol_to_str(buffer, sizeof(buffer), mpol);
1348
1349         seq_printf(seq, ",mpol=%s", buffer);
1350 }
1351
1352 static struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1353 {
1354         struct mempolicy *mpol = NULL;
1355         if (sbinfo->mpol) {
1356                 spin_lock(&sbinfo->stat_lock);  /* prevent replace/use races */
1357                 mpol = sbinfo->mpol;
1358                 mpol_get(mpol);
1359                 spin_unlock(&sbinfo->stat_lock);
1360         }
1361         return mpol;
1362 }
1363 #else /* !CONFIG_NUMA || !CONFIG_TMPFS */
1364 static inline void shmem_show_mpol(struct seq_file *seq, struct mempolicy *mpol)
1365 {
1366 }
1367 static inline struct mempolicy *shmem_get_sbmpol(struct shmem_sb_info *sbinfo)
1368 {
1369         return NULL;
1370 }
1371 #endif /* CONFIG_NUMA && CONFIG_TMPFS */
1372 #ifndef CONFIG_NUMA
1373 #define vm_policy vm_private_data
1374 #endif
1375
1376 static void shmem_pseudo_vma_init(struct vm_area_struct *vma,
1377                 struct shmem_inode_info *info, pgoff_t index)
1378 {
1379         /* Create a pseudo vma that just contains the policy */
1380         vma->vm_start = 0;
1381         /* Bias interleave by inode number to distribute better across nodes */
1382         vma->vm_pgoff = index + info->vfs_inode.i_ino;
1383         vma->vm_ops = NULL;
1384         vma->vm_policy = mpol_shared_policy_lookup(&info->policy, index);
1385 }
1386
1387 static void shmem_pseudo_vma_destroy(struct vm_area_struct *vma)
1388 {
1389         /* Drop reference taken by mpol_shared_policy_lookup() */
1390         mpol_cond_put(vma->vm_policy);
1391 }
1392
1393 static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
1394                         struct shmem_inode_info *info, pgoff_t index)
1395 {
1396         struct vm_area_struct pvma;
1397         struct page *page;
1398
1399         shmem_pseudo_vma_init(&pvma, info, index);
1400         page = swapin_readahead(swap, gfp, &pvma, 0);
1401         shmem_pseudo_vma_destroy(&pvma);
1402
1403         return page;
1404 }
1405
1406 static struct page *shmem_alloc_hugepage(gfp_t gfp,
1407                 struct shmem_inode_info *info, pgoff_t index)
1408 {
1409         struct vm_area_struct pvma;
1410         struct inode *inode = &info->vfs_inode;
1411         struct address_space *mapping = inode->i_mapping;
1412         pgoff_t idx, hindex;
1413         void __rcu **results;
1414         struct page *page;
1415
1416         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1417                 return NULL;
1418
1419         hindex = round_down(index, HPAGE_PMD_NR);
1420         rcu_read_lock();
1421         if (radix_tree_gang_lookup_slot(&mapping->page_tree, &results, &idx,
1422                                 hindex, 1) && idx < hindex + HPAGE_PMD_NR) {
1423                 rcu_read_unlock();
1424                 return NULL;
1425         }
1426         rcu_read_unlock();
1427
1428         shmem_pseudo_vma_init(&pvma, info, hindex);
1429         page = alloc_pages_vma(gfp | __GFP_COMP | __GFP_NORETRY | __GFP_NOWARN,
1430                         HPAGE_PMD_ORDER, &pvma, 0, numa_node_id(), true);
1431         shmem_pseudo_vma_destroy(&pvma);
1432         if (page)
1433                 prep_transhuge_page(page);
1434         return page;
1435 }
1436
1437 static struct page *shmem_alloc_page(gfp_t gfp,
1438                         struct shmem_inode_info *info, pgoff_t index)
1439 {
1440         struct vm_area_struct pvma;
1441         struct page *page;
1442
1443         shmem_pseudo_vma_init(&pvma, info, index);
1444         page = alloc_page_vma(gfp, &pvma, 0);
1445         shmem_pseudo_vma_destroy(&pvma);
1446
1447         return page;
1448 }
1449
1450 static struct page *shmem_alloc_and_acct_page(gfp_t gfp,
1451                 struct shmem_inode_info *info, struct shmem_sb_info *sbinfo,
1452                 pgoff_t index, bool huge)
1453 {
1454         struct page *page;
1455         int nr;
1456         int err = -ENOSPC;
1457
1458         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
1459                 huge = false;
1460         nr = huge ? HPAGE_PMD_NR : 1;
1461
1462         if (shmem_acct_block(info->flags, nr))
1463                 goto failed;
1464         if (sbinfo->max_blocks) {
1465                 if (percpu_counter_compare(&sbinfo->used_blocks,
1466                                         sbinfo->max_blocks - nr) > 0)
1467                         goto unacct;
1468                 percpu_counter_add(&sbinfo->used_blocks, nr);
1469         }
1470
1471         if (huge)
1472                 page = shmem_alloc_hugepage(gfp, info, index);
1473         else
1474                 page = shmem_alloc_page(gfp, info, index);
1475         if (page) {
1476                 __SetPageLocked(page);
1477                 __SetPageSwapBacked(page);
1478                 return page;
1479         }
1480
1481         err = -ENOMEM;
1482         if (sbinfo->max_blocks)
1483                 percpu_counter_add(&sbinfo->used_blocks, -nr);
1484 unacct:
1485         shmem_unacct_blocks(info->flags, nr);
1486 failed:
1487         return ERR_PTR(err);
1488 }
1489
1490 /*
1491  * When a page is moved from swapcache to shmem filecache (either by the
1492  * usual swapin of shmem_getpage_gfp(), or by the less common swapoff of
1493  * shmem_unuse_inode()), it may have been read in earlier from swap, in
1494  * ignorance of the mapping it belongs to.  If that mapping has special
1495  * constraints (like the gma500 GEM driver, which requires RAM below 4GB),
1496  * we may need to copy to a suitable page before moving to filecache.
1497  *
1498  * In a future release, this may well be extended to respect cpuset and
1499  * NUMA mempolicy, and applied also to anonymous pages in do_swap_page();
1500  * but for now it is a simple matter of zone.
1501  */
1502 static bool shmem_should_replace_page(struct page *page, gfp_t gfp)
1503 {
1504         return page_zonenum(page) > gfp_zone(gfp);
1505 }
1506
1507 static int shmem_replace_page(struct page **pagep, gfp_t gfp,
1508                                 struct shmem_inode_info *info, pgoff_t index)
1509 {
1510         struct page *oldpage, *newpage;
1511         struct address_space *swap_mapping;
1512         pgoff_t swap_index;
1513         int error;
1514
1515         oldpage = *pagep;
1516         swap_index = page_private(oldpage);
1517         swap_mapping = page_mapping(oldpage);
1518
1519         /*
1520          * We have arrived here because our zones are constrained, so don't
1521          * limit chance of success by further cpuset and node constraints.
1522          */
1523         gfp &= ~GFP_CONSTRAINT_MASK;
1524         newpage = shmem_alloc_page(gfp, info, index);
1525         if (!newpage)
1526                 return -ENOMEM;
1527
1528         get_page(newpage);
1529         copy_highpage(newpage, oldpage);
1530         flush_dcache_page(newpage);
1531
1532         __SetPageLocked(newpage);
1533         __SetPageSwapBacked(newpage);
1534         SetPageUptodate(newpage);
1535         set_page_private(newpage, swap_index);
1536         SetPageSwapCache(newpage);
1537
1538         /*
1539          * Our caller will very soon move newpage out of swapcache, but it's
1540          * a nice clean interface for us to replace oldpage by newpage there.
1541          */
1542         spin_lock_irq(&swap_mapping->tree_lock);
1543         error = shmem_radix_tree_replace(swap_mapping, swap_index, oldpage,
1544                                                                    newpage);
1545         if (!error) {
1546                 __inc_node_page_state(newpage, NR_FILE_PAGES);
1547                 __dec_node_page_state(oldpage, NR_FILE_PAGES);
1548         }
1549         spin_unlock_irq(&swap_mapping->tree_lock);
1550
1551         if (unlikely(error)) {
1552                 /*
1553                  * Is this possible?  I think not, now that our callers check
1554                  * both PageSwapCache and page_private after getting page lock;
1555                  * but be defensive.  Reverse old to newpage for clear and free.
1556                  */
1557                 oldpage = newpage;
1558         } else {
1559                 mem_cgroup_migrate(oldpage, newpage);
1560                 lru_cache_add_anon(newpage);
1561                 *pagep = newpage;
1562         }
1563
1564         ClearPageSwapCache(oldpage);
1565         set_page_private(oldpage, 0);
1566
1567         unlock_page(oldpage);
1568         put_page(oldpage);
1569         put_page(oldpage);
1570         return error;
1571 }
1572
1573 /*
1574  * shmem_getpage_gfp - find page in cache, or get from swap, or allocate
1575  *
1576  * If we allocate a new one we do not mark it dirty. That's up to the
1577  * vm. If we swap it in we mark it dirty since we also free the swap
1578  * entry since a page cannot live in both the swap and page cache.
1579  *
1580  * fault_mm and fault_type are only supplied by shmem_fault:
1581  * otherwise they are NULL.
1582  */
1583 static int shmem_getpage_gfp(struct inode *inode, pgoff_t index,
1584         struct page **pagep, enum sgp_type sgp, gfp_t gfp,
1585         struct vm_area_struct *vma, struct vm_fault *vmf, int *fault_type)
1586 {
1587         struct address_space *mapping = inode->i_mapping;
1588         struct shmem_inode_info *info = SHMEM_I(inode);
1589         struct shmem_sb_info *sbinfo;
1590         struct mm_struct *charge_mm;
1591         struct mem_cgroup *memcg;
1592         struct page *page;
1593         swp_entry_t swap;
1594         enum sgp_type sgp_huge = sgp;
1595         pgoff_t hindex = index;
1596         int error;
1597         int once = 0;
1598         int alloced = 0;
1599
1600         if (index > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
1601                 return -EFBIG;
1602         if (sgp == SGP_NOHUGE || sgp == SGP_HUGE)
1603                 sgp = SGP_CACHE;
1604 repeat:
1605         swap.val = 0;
1606         page = find_lock_entry(mapping, index);
1607         if (radix_tree_exceptional_entry(page)) {
1608                 swap = radix_to_swp_entry(page);
1609                 page = NULL;
1610         }
1611
1612         if (sgp <= SGP_CACHE &&
1613             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1614                 error = -EINVAL;
1615                 goto unlock;
1616         }
1617
1618         if (page && sgp == SGP_WRITE)
1619                 mark_page_accessed(page);
1620
1621         /* fallocated page? */
1622         if (page && !PageUptodate(page)) {
1623                 if (sgp != SGP_READ)
1624                         goto clear;
1625                 unlock_page(page);
1626                 put_page(page);
1627                 page = NULL;
1628         }
1629         if (page || (sgp == SGP_READ && !swap.val)) {
1630                 *pagep = page;
1631                 return 0;
1632         }
1633
1634         /*
1635          * Fast cache lookup did not find it:
1636          * bring it back from swap or allocate.
1637          */
1638         sbinfo = SHMEM_SB(inode->i_sb);
1639         charge_mm = vma ? vma->vm_mm : current->mm;
1640
1641         if (swap.val) {
1642                 /* Look it up and read it in.. */
1643                 page = lookup_swap_cache(swap);
1644                 if (!page) {
1645                         /* Or update major stats only when swapin succeeds?? */
1646                         if (fault_type) {
1647                                 *fault_type |= VM_FAULT_MAJOR;
1648                                 count_vm_event(PGMAJFAULT);
1649                                 count_memcg_event_mm(charge_mm, PGMAJFAULT);
1650                         }
1651                         /* Here we actually start the io */
1652                         page = shmem_swapin(swap, gfp, info, index);
1653                         if (!page) {
1654                                 error = -ENOMEM;
1655                                 goto failed;
1656                         }
1657                 }
1658
1659                 /* We have to do this with page locked to prevent races */
1660                 lock_page(page);
1661                 if (!PageSwapCache(page) || page_private(page) != swap.val ||
1662                     !shmem_confirm_swap(mapping, index, swap)) {
1663                         error = -EEXIST;        /* try again */
1664                         goto unlock;
1665                 }
1666                 if (!PageUptodate(page)) {
1667                         error = -EIO;
1668                         goto failed;
1669                 }
1670                 wait_on_page_writeback(page);
1671
1672                 if (shmem_should_replace_page(page, gfp)) {
1673                         error = shmem_replace_page(&page, gfp, info, index);
1674                         if (error)
1675                                 goto failed;
1676                 }
1677
1678                 error = mem_cgroup_try_charge(page, charge_mm, gfp, &memcg,
1679                                 false);
1680                 if (!error) {
1681                         error = shmem_add_to_page_cache(page, mapping, index,
1682                                                 swp_to_radix_entry(swap));
1683                         /*
1684                          * We already confirmed swap under page lock, and make
1685                          * no memory allocation here, so usually no possibility
1686                          * of error; but free_swap_and_cache() only trylocks a
1687                          * page, so it is just possible that the entry has been
1688                          * truncated or holepunched since swap was confirmed.
1689                          * shmem_undo_range() will have done some of the
1690                          * unaccounting, now delete_from_swap_cache() will do
1691                          * the rest.
1692                          * Reset swap.val? No, leave it so "failed" goes back to
1693                          * "repeat": reading a hole and writing should succeed.
1694                          */
1695                         if (error) {
1696                                 mem_cgroup_cancel_charge(page, memcg, false);
1697                                 delete_from_swap_cache(page);
1698                         }
1699                 }
1700                 if (error)
1701                         goto failed;
1702
1703                 mem_cgroup_commit_charge(page, memcg, true, false);
1704
1705                 spin_lock_irq(&info->lock);
1706                 info->swapped--;
1707                 shmem_recalc_inode(inode);
1708                 spin_unlock_irq(&info->lock);
1709
1710                 if (sgp == SGP_WRITE)
1711                         mark_page_accessed(page);
1712
1713                 delete_from_swap_cache(page);
1714                 set_page_dirty(page);
1715                 swap_free(swap);
1716
1717         } else {
1718                 if (vma && userfaultfd_missing(vma)) {
1719                         *fault_type = handle_userfault(vmf, VM_UFFD_MISSING);
1720                         return 0;
1721                 }
1722
1723                 /* shmem_symlink() */
1724                 if (mapping->a_ops != &shmem_aops)
1725                         goto alloc_nohuge;
1726                 if (shmem_huge == SHMEM_HUGE_DENY || sgp_huge == SGP_NOHUGE)
1727                         goto alloc_nohuge;
1728                 if (shmem_huge == SHMEM_HUGE_FORCE)
1729                         goto alloc_huge;
1730                 switch (sbinfo->huge) {
1731                         loff_t i_size;
1732                         pgoff_t off;
1733                 case SHMEM_HUGE_NEVER:
1734                         goto alloc_nohuge;
1735                 case SHMEM_HUGE_WITHIN_SIZE:
1736                         off = round_up(index, HPAGE_PMD_NR);
1737                         i_size = round_up(i_size_read(inode), PAGE_SIZE);
1738                         if (i_size >= HPAGE_PMD_SIZE &&
1739                                         i_size >> PAGE_SHIFT >= off)
1740                                 goto alloc_huge;
1741                         /* fallthrough */
1742                 case SHMEM_HUGE_ADVISE:
1743                         if (sgp_huge == SGP_HUGE)
1744                                 goto alloc_huge;
1745                         /* TODO: implement fadvise() hints */
1746                         goto alloc_nohuge;
1747                 }
1748
1749 alloc_huge:
1750                 page = shmem_alloc_and_acct_page(gfp, info, sbinfo,
1751                                 index, true);
1752                 if (IS_ERR(page)) {
1753 alloc_nohuge:           page = shmem_alloc_and_acct_page(gfp, info, sbinfo,
1754                                         index, false);
1755                 }
1756                 if (IS_ERR(page)) {
1757                         int retry = 5;
1758                         error = PTR_ERR(page);
1759                         page = NULL;
1760                         if (error != -ENOSPC)
1761                                 goto failed;
1762                         /*
1763                          * Try to reclaim some spece by splitting a huge page
1764                          * beyond i_size on the filesystem.
1765                          */
1766                         while (retry--) {
1767                                 int ret;
1768                                 ret = shmem_unused_huge_shrink(sbinfo, NULL, 1);
1769                                 if (ret == SHRINK_STOP)
1770                                         break;
1771                                 if (ret)
1772                                         goto alloc_nohuge;
1773                         }
1774                         goto failed;
1775                 }
1776
1777                 if (PageTransHuge(page))
1778                         hindex = round_down(index, HPAGE_PMD_NR);
1779                 else
1780                         hindex = index;
1781
1782                 if (sgp == SGP_WRITE)
1783                         __SetPageReferenced(page);
1784
1785                 error = mem_cgroup_try_charge(page, charge_mm, gfp, &memcg,
1786                                 PageTransHuge(page));
1787                 if (error)
1788                         goto unacct;
1789                 error = radix_tree_maybe_preload_order(gfp & GFP_RECLAIM_MASK,
1790                                 compound_order(page));
1791                 if (!error) {
1792                         error = shmem_add_to_page_cache(page, mapping, hindex,
1793                                                         NULL);
1794                         radix_tree_preload_end();
1795                 }
1796                 if (error) {
1797                         mem_cgroup_cancel_charge(page, memcg,
1798                                         PageTransHuge(page));
1799                         goto unacct;
1800                 }
1801                 mem_cgroup_commit_charge(page, memcg, false,
1802                                 PageTransHuge(page));
1803                 lru_cache_add_anon(page);
1804
1805                 spin_lock_irq(&info->lock);
1806                 info->alloced += 1 << compound_order(page);
1807                 inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
1808                 shmem_recalc_inode(inode);
1809                 spin_unlock_irq(&info->lock);
1810                 alloced = true;
1811
1812                 if (PageTransHuge(page) &&
1813                                 DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE) <
1814                                 hindex + HPAGE_PMD_NR - 1) {
1815                         /*
1816                          * Part of the huge page is beyond i_size: subject
1817                          * to shrink under memory pressure.
1818                          */
1819                         spin_lock(&sbinfo->shrinklist_lock);
1820                         if (list_empty(&info->shrinklist)) {
1821                                 list_add_tail(&info->shrinklist,
1822                                                 &sbinfo->shrinklist);
1823                                 sbinfo->shrinklist_len++;
1824                         }
1825                         spin_unlock(&sbinfo->shrinklist_lock);
1826                 }
1827
1828                 /*
1829                  * Let SGP_FALLOC use the SGP_WRITE optimization on a new page.
1830                  */
1831                 if (sgp == SGP_FALLOC)
1832                         sgp = SGP_WRITE;
1833 clear:
1834                 /*
1835                  * Let SGP_WRITE caller clear ends if write does not fill page;
1836                  * but SGP_FALLOC on a page fallocated earlier must initialize
1837                  * it now, lest undo on failure cancel our earlier guarantee.
1838                  */
1839                 if (sgp != SGP_WRITE && !PageUptodate(page)) {
1840                         struct page *head = compound_head(page);
1841                         int i;
1842
1843                         for (i = 0; i < (1 << compound_order(head)); i++) {
1844                                 clear_highpage(head + i);
1845                                 flush_dcache_page(head + i);
1846                         }
1847                         SetPageUptodate(head);
1848                 }
1849         }
1850
1851         /* Perhaps the file has been truncated since we checked */
1852         if (sgp <= SGP_CACHE &&
1853             ((loff_t)index << PAGE_SHIFT) >= i_size_read(inode)) {
1854                 if (alloced) {
1855                         ClearPageDirty(page);
1856                         delete_from_page_cache(page);
1857                         spin_lock_irq(&info->lock);
1858                         shmem_recalc_inode(inode);
1859                         spin_unlock_irq(&info->lock);
1860                 }
1861                 error = -EINVAL;
1862                 goto unlock;
1863         }
1864         *pagep = page + index - hindex;
1865         return 0;
1866
1867         /*
1868          * Error recovery.
1869          */
1870 unacct:
1871         if (sbinfo->max_blocks)
1872                 percpu_counter_sub(&sbinfo->used_blocks,
1873                                 1 << compound_order(page));
1874         shmem_unacct_blocks(info->flags, 1 << compound_order(page));
1875
1876         if (PageTransHuge(page)) {
1877                 unlock_page(page);
1878                 put_page(page);
1879                 goto alloc_nohuge;
1880         }
1881 failed:
1882         if (swap.val && !shmem_confirm_swap(mapping, index, swap))
1883                 error = -EEXIST;
1884 unlock:
1885         if (page) {
1886                 unlock_page(page);
1887                 put_page(page);
1888         }
1889         if (error == -ENOSPC && !once++) {
1890                 spin_lock_irq(&info->lock);
1891                 shmem_recalc_inode(inode);
1892                 spin_unlock_irq(&info->lock);
1893                 goto repeat;
1894         }
1895         if (error == -EEXIST)   /* from above or from radix_tree_insert */
1896                 goto repeat;
1897         return error;
1898 }
1899
1900 /*
1901  * This is like autoremove_wake_function, but it removes the wait queue
1902  * entry unconditionally - even if something else had already woken the
1903  * target.
1904  */
1905 static int synchronous_wake_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
1906 {
1907         int ret = default_wake_function(wait, mode, sync, key);
1908         list_del_init(&wait->entry);
1909         return ret;
1910 }
1911
1912 static int shmem_fault(struct vm_fault *vmf)
1913 {
1914         struct vm_area_struct *vma = vmf->vma;
1915         struct inode *inode = file_inode(vma->vm_file);
1916         gfp_t gfp = mapping_gfp_mask(inode->i_mapping);
1917         enum sgp_type sgp;
1918         int error;
1919         int ret = VM_FAULT_LOCKED;
1920
1921         /*
1922          * Trinity finds that probing a hole which tmpfs is punching can
1923          * prevent the hole-punch from ever completing: which in turn
1924          * locks writers out with its hold on i_mutex.  So refrain from
1925          * faulting pages into the hole while it's being punched.  Although
1926          * shmem_undo_range() does remove the additions, it may be unable to
1927          * keep up, as each new page needs its own unmap_mapping_range() call,
1928          * and the i_mmap tree grows ever slower to scan if new vmas are added.
1929          *
1930          * It does not matter if we sometimes reach this check just before the
1931          * hole-punch begins, so that one fault then races with the punch:
1932          * we just need to make racing faults a rare case.
1933          *
1934          * The implementation below would be much simpler if we just used a
1935          * standard mutex or completion: but we cannot take i_mutex in fault,
1936          * and bloating every shmem inode for this unlikely case would be sad.
1937          */
1938         if (unlikely(inode->i_private)) {
1939                 struct shmem_falloc *shmem_falloc;
1940
1941                 spin_lock(&inode->i_lock);
1942                 shmem_falloc = inode->i_private;
1943                 if (shmem_falloc &&
1944                     shmem_falloc->waitq &&
1945                     vmf->pgoff >= shmem_falloc->start &&
1946                     vmf->pgoff < shmem_falloc->next) {
1947                         wait_queue_head_t *shmem_falloc_waitq;
1948                         DEFINE_WAIT_FUNC(shmem_fault_wait, synchronous_wake_function);
1949
1950                         ret = VM_FAULT_NOPAGE;
1951                         if ((vmf->flags & FAULT_FLAG_ALLOW_RETRY) &&
1952                            !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT)) {
1953                                 /* It's polite to up mmap_sem if we can */
1954                                 up_read(&vma->vm_mm->mmap_sem);
1955                                 ret = VM_FAULT_RETRY;
1956                         }
1957
1958                         shmem_falloc_waitq = shmem_falloc->waitq;
1959                         prepare_to_wait(shmem_falloc_waitq, &shmem_fault_wait,
1960                                         TASK_UNINTERRUPTIBLE);
1961                         spin_unlock(&inode->i_lock);
1962                         schedule();
1963
1964                         /*
1965                          * shmem_falloc_waitq points into the shmem_fallocate()
1966                          * stack of the hole-punching task: shmem_falloc_waitq
1967                          * is usually invalid by the time we reach here, but
1968                          * finish_wait() does not dereference it in that case;
1969                          * though i_lock needed lest racing with wake_up_all().
1970                          */
1971                         spin_lock(&inode->i_lock);
1972                         finish_wait(shmem_falloc_waitq, &shmem_fault_wait);
1973                         spin_unlock(&inode->i_lock);
1974                         return ret;
1975                 }
1976                 spin_unlock(&inode->i_lock);
1977         }
1978
1979         sgp = SGP_CACHE;
1980
1981         if ((vma->vm_flags & VM_NOHUGEPAGE) ||
1982             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
1983                 sgp = SGP_NOHUGE;
1984         else if (vma->vm_flags & VM_HUGEPAGE)
1985                 sgp = SGP_HUGE;
1986
1987         error = shmem_getpage_gfp(inode, vmf->pgoff, &vmf->page, sgp,
1988                                   gfp, vma, vmf, &ret);
1989         if (error)
1990                 return ((error == -ENOMEM) ? VM_FAULT_OOM : VM_FAULT_SIGBUS);
1991         return ret;
1992 }
1993
1994 unsigned long shmem_get_unmapped_area(struct file *file,
1995                                       unsigned long uaddr, unsigned long len,
1996                                       unsigned long pgoff, unsigned long flags)
1997 {
1998         unsigned long (*get_area)(struct file *,
1999                 unsigned long, unsigned long, unsigned long, unsigned long);
2000         unsigned long addr;
2001         unsigned long offset;
2002         unsigned long inflated_len;
2003         unsigned long inflated_addr;
2004         unsigned long inflated_offset;
2005
2006         if (len > TASK_SIZE)
2007                 return -ENOMEM;
2008
2009         get_area = current->mm->get_unmapped_area;
2010         addr = get_area(file, uaddr, len, pgoff, flags);
2011
2012         if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
2013                 return addr;
2014         if (IS_ERR_VALUE(addr))
2015                 return addr;
2016         if (addr & ~PAGE_MASK)
2017                 return addr;
2018         if (addr > TASK_SIZE - len)
2019                 return addr;
2020
2021         if (shmem_huge == SHMEM_HUGE_DENY)
2022                 return addr;
2023         if (len < HPAGE_PMD_SIZE)
2024                 return addr;
2025         if (flags & MAP_FIXED)
2026                 return addr;
2027         /*
2028          * Our priority is to support MAP_SHARED mapped hugely;
2029          * and support MAP_PRIVATE mapped hugely too, until it is COWed.
2030          * But if caller specified an address hint, respect that as before.
2031          */
2032         if (uaddr)
2033                 return addr;
2034
2035         if (shmem_huge != SHMEM_HUGE_FORCE) {
2036                 struct super_block *sb;
2037
2038                 if (file) {
2039                         VM_BUG_ON(file->f_op != &shmem_file_operations);
2040                         sb = file_inode(file)->i_sb;
2041                 } else {
2042                         /*
2043                          * Called directly from mm/mmap.c, or drivers/char/mem.c
2044                          * for "/dev/zero", to create a shared anonymous object.
2045                          */
2046                         if (IS_ERR(shm_mnt))
2047                                 return addr;
2048                         sb = shm_mnt->mnt_sb;
2049                 }
2050                 if (SHMEM_SB(sb)->huge == SHMEM_HUGE_NEVER)
2051                         return addr;
2052         }
2053
2054         offset = (pgoff << PAGE_SHIFT) & (HPAGE_PMD_SIZE-1);
2055         if (offset && offset + len < 2 * HPAGE_PMD_SIZE)
2056                 return addr;
2057         if ((addr & (HPAGE_PMD_SIZE-1)) == offset)
2058                 return addr;
2059
2060         inflated_len = len + HPAGE_PMD_SIZE - PAGE_SIZE;
2061         if (inflated_len > TASK_SIZE)
2062                 return addr;
2063         if (inflated_len < len)
2064                 return addr;
2065
2066         inflated_addr = get_area(NULL, 0, inflated_len, 0, flags);
2067         if (IS_ERR_VALUE(inflated_addr))
2068                 return addr;
2069         if (inflated_addr & ~PAGE_MASK)
2070                 return addr;
2071
2072         inflated_offset = inflated_addr & (HPAGE_PMD_SIZE-1);
2073         inflated_addr += offset - inflated_offset;
2074         if (inflated_offset > offset)
2075                 inflated_addr += HPAGE_PMD_SIZE;
2076
2077         if (inflated_addr > TASK_SIZE - len)
2078                 return addr;
2079         return inflated_addr;
2080 }
2081
2082 #ifdef CONFIG_NUMA
2083 static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
2084 {
2085         struct inode *inode = file_inode(vma->vm_file);
2086         return mpol_set_shared_policy(&SHMEM_I(inode)->policy, vma, mpol);
2087 }
2088
2089 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
2090                                           unsigned long addr)
2091 {
2092         struct inode *inode = file_inode(vma->vm_file);
2093         pgoff_t index;
2094
2095         index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2096         return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
2097 }
2098 #endif
2099
2100 int shmem_lock(struct file *file, int lock, struct user_struct *user)
2101 {
2102         struct inode *inode = file_inode(file);
2103         struct shmem_inode_info *info = SHMEM_I(inode);
2104         int retval = -ENOMEM;
2105
2106         spin_lock_irq(&info->lock);
2107         if (lock && !(info->flags & VM_LOCKED)) {
2108                 if (!user_shm_lock(inode->i_size, user))
2109                         goto out_nomem;
2110                 info->flags |= VM_LOCKED;
2111                 mapping_set_unevictable(file->f_mapping);
2112         }
2113         if (!lock && (info->flags & VM_LOCKED) && user) {
2114                 user_shm_unlock(inode->i_size, user);
2115                 info->flags &= ~VM_LOCKED;
2116                 mapping_clear_unevictable(file->f_mapping);
2117         }
2118         retval = 0;
2119
2120 out_nomem:
2121         spin_unlock_irq(&info->lock);
2122         return retval;
2123 }
2124
2125 static int shmem_mmap(struct file *file, struct vm_area_struct *vma)
2126 {
2127         file_accessed(file);
2128         vma->vm_ops = &shmem_vm_ops;
2129         if (IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE) &&
2130                         ((vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK) <
2131                         (vma->vm_end & HPAGE_PMD_MASK)) {
2132                 khugepaged_enter(vma, vma->vm_flags);
2133         }
2134         return 0;
2135 }
2136
2137 static struct inode *shmem_get_inode(struct super_block *sb, const struct inode *dir,
2138                                      umode_t mode, dev_t dev, unsigned long flags)
2139 {
2140         struct inode *inode;
2141         struct shmem_inode_info *info;
2142         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
2143
2144         if (shmem_reserve_inode(sb))
2145                 return NULL;
2146
2147         inode = new_inode(sb);
2148         if (inode) {
2149                 inode->i_ino = get_next_ino();
2150                 inode_init_owner(inode, dir, mode);
2151                 inode->i_blocks = 0;
2152                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
2153                 inode->i_generation = get_seconds();
2154                 info = SHMEM_I(inode);
2155                 memset(info, 0, (char *)inode - (char *)info);
2156                 spin_lock_init(&info->lock);
2157                 info->seals = F_SEAL_SEAL;
2158                 info->flags = flags & VM_NORESERVE;
2159                 INIT_LIST_HEAD(&info->shrinklist);
2160                 INIT_LIST_HEAD(&info->swaplist);
2161                 simple_xattrs_init(&info->xattrs);
2162                 cache_no_acl(inode);
2163
2164                 switch (mode & S_IFMT) {
2165                 default:
2166                         inode->i_op = &shmem_special_inode_operations;
2167                         init_special_inode(inode, mode, dev);
2168                         break;
2169                 case S_IFREG:
2170                         inode->i_mapping->a_ops = &shmem_aops;
2171                         inode->i_op = &shmem_inode_operations;
2172                         inode->i_fop = &shmem_file_operations;
2173                         mpol_shared_policy_init(&info->policy,
2174                                                  shmem_get_sbmpol(sbinfo));
2175                         break;
2176                 case S_IFDIR:
2177                         inc_nlink(inode);
2178                         /* Some things misbehave if size == 0 on a directory */
2179                         inode->i_size = 2 * BOGO_DIRENT_SIZE;
2180                         inode->i_op = &shmem_dir_inode_operations;
2181                         inode->i_fop = &simple_dir_operations;
2182                         break;
2183                 case S_IFLNK:
2184                         /*
2185                          * Must not load anything in the rbtree,
2186                          * mpol_free_shared_policy will not be called.
2187                          */
2188                         mpol_shared_policy_init(&info->policy, NULL);
2189                         break;
2190                 }
2191         } else
2192                 shmem_free_inode(sb);
2193         return inode;
2194 }
2195
2196 bool shmem_mapping(struct address_space *mapping)
2197 {
2198         return mapping->a_ops == &shmem_aops;
2199 }
2200
2201 int shmem_mcopy_atomic_pte(struct mm_struct *dst_mm,
2202                            pmd_t *dst_pmd,
2203                            struct vm_area_struct *dst_vma,
2204                            unsigned long dst_addr,
2205                            unsigned long src_addr,
2206                            struct page **pagep)
2207 {
2208         struct inode *inode = file_inode(dst_vma->vm_file);
2209         struct shmem_inode_info *info = SHMEM_I(inode);
2210         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2211         struct address_space *mapping = inode->i_mapping;
2212         gfp_t gfp = mapping_gfp_mask(mapping);
2213         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
2214         struct mem_cgroup *memcg;
2215         spinlock_t *ptl;
2216         void *page_kaddr;
2217         struct page *page;
2218         pte_t _dst_pte, *dst_pte;
2219         int ret;
2220
2221         ret = -ENOMEM;
2222         if (shmem_acct_block(info->flags, 1))
2223                 goto out;
2224         if (sbinfo->max_blocks) {
2225                 if (percpu_counter_compare(&sbinfo->used_blocks,
2226                                            sbinfo->max_blocks) >= 0)
2227                         goto out_unacct_blocks;
2228                 percpu_counter_inc(&sbinfo->used_blocks);
2229         }
2230
2231         if (!*pagep) {
2232                 page = shmem_alloc_page(gfp, info, pgoff);
2233                 if (!page)
2234                         goto out_dec_used_blocks;
2235
2236                 page_kaddr = kmap_atomic(page);
2237                 ret = copy_from_user(page_kaddr, (const void __user *)src_addr,
2238                                      PAGE_SIZE);
2239                 kunmap_atomic(page_kaddr);
2240
2241                 /* fallback to copy_from_user outside mmap_sem */
2242                 if (unlikely(ret)) {
2243                         *pagep = page;
2244                         if (sbinfo->max_blocks)
2245                                 percpu_counter_add(&sbinfo->used_blocks, -1);
2246                         shmem_unacct_blocks(info->flags, 1);
2247                         /* don't free the page */
2248                         return -EFAULT;
2249                 }
2250         } else {
2251                 page = *pagep;
2252                 *pagep = NULL;
2253         }
2254
2255         VM_BUG_ON(PageLocked(page) || PageSwapBacked(page));
2256         __SetPageLocked(page);
2257         __SetPageSwapBacked(page);
2258         __SetPageUptodate(page);
2259
2260         ret = mem_cgroup_try_charge(page, dst_mm, gfp, &memcg, false);
2261         if (ret)
2262                 goto out_release;
2263
2264         ret = radix_tree_maybe_preload(gfp & GFP_RECLAIM_MASK);
2265         if (!ret) {
2266                 ret = shmem_add_to_page_cache(page, mapping, pgoff, NULL);
2267                 radix_tree_preload_end();
2268         }
2269         if (ret)
2270                 goto out_release_uncharge;
2271
2272         mem_cgroup_commit_charge(page, memcg, false, false);
2273
2274         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
2275         if (dst_vma->vm_flags & VM_WRITE)
2276                 _dst_pte = pte_mkwrite(pte_mkdirty(_dst_pte));
2277
2278         ret = -EEXIST;
2279         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
2280         if (!pte_none(*dst_pte))
2281                 goto out_release_uncharge_unlock;
2282
2283         lru_cache_add_anon(page);
2284
2285         spin_lock(&info->lock);
2286         info->alloced++;
2287         inode->i_blocks += BLOCKS_PER_PAGE;
2288         shmem_recalc_inode(inode);
2289         spin_unlock(&info->lock);
2290
2291         inc_mm_counter(dst_mm, mm_counter_file(page));
2292         page_add_file_rmap(page, false);
2293         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
2294
2295         /* No need to invalidate - it was non-present before */
2296         update_mmu_cache(dst_vma, dst_addr, dst_pte);
2297         unlock_page(page);
2298         pte_unmap_unlock(dst_pte, ptl);
2299         ret = 0;
2300 out:
2301         return ret;
2302 out_release_uncharge_unlock:
2303         pte_unmap_unlock(dst_pte, ptl);
2304 out_release_uncharge:
2305         mem_cgroup_cancel_charge(page, memcg, false);
2306 out_release:
2307         unlock_page(page);
2308         put_page(page);
2309 out_dec_used_blocks:
2310         if (sbinfo->max_blocks)
2311                 percpu_counter_add(&sbinfo->used_blocks, -1);
2312 out_unacct_blocks:
2313         shmem_unacct_blocks(info->flags, 1);
2314         goto out;
2315 }
2316
2317 #ifdef CONFIG_TMPFS
2318 static const struct inode_operations shmem_symlink_inode_operations;
2319 static const struct inode_operations shmem_short_symlink_operations;
2320
2321 #ifdef CONFIG_TMPFS_XATTR
2322 static int shmem_initxattrs(struct inode *, const struct xattr *, void *);
2323 #else
2324 #define shmem_initxattrs NULL
2325 #endif
2326
2327 static int
2328 shmem_write_begin(struct file *file, struct address_space *mapping,
2329                         loff_t pos, unsigned len, unsigned flags,
2330                         struct page **pagep, void **fsdata)
2331 {
2332         struct inode *inode = mapping->host;
2333         struct shmem_inode_info *info = SHMEM_I(inode);
2334         pgoff_t index = pos >> PAGE_SHIFT;
2335
2336         /* i_mutex is held by caller */
2337         if (unlikely(info->seals & (F_SEAL_WRITE | F_SEAL_GROW))) {
2338                 if (info->seals & F_SEAL_WRITE)
2339                         return -EPERM;
2340                 if ((info->seals & F_SEAL_GROW) && pos + len > inode->i_size)
2341                         return -EPERM;
2342         }
2343
2344         return shmem_getpage(inode, index, pagep, SGP_WRITE);
2345 }
2346
2347 static int
2348 shmem_write_end(struct file *file, struct address_space *mapping,
2349                         loff_t pos, unsigned len, unsigned copied,
2350                         struct page *page, void *fsdata)
2351 {
2352         struct inode *inode = mapping->host;
2353
2354         if (pos + copied > inode->i_size)
2355                 i_size_write(inode, pos + copied);
2356
2357         if (!PageUptodate(page)) {
2358                 struct page *head = compound_head(page);
2359                 if (PageTransCompound(page)) {
2360                         int i;
2361
2362                         for (i = 0; i < HPAGE_PMD_NR; i++) {
2363                                 if (head + i == page)
2364                                         continue;
2365                                 clear_highpage(head + i);
2366                                 flush_dcache_page(head + i);
2367                         }
2368                 }
2369                 if (copied < PAGE_SIZE) {
2370                         unsigned from = pos & (PAGE_SIZE - 1);
2371                         zero_user_segments(page, 0, from,
2372                                         from + copied, PAGE_SIZE);
2373                 }
2374                 SetPageUptodate(head);
2375         }
2376         set_page_dirty(page);
2377         unlock_page(page);
2378         put_page(page);
2379
2380         return copied;
2381 }
2382
2383 static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
2384 {
2385         struct file *file = iocb->ki_filp;
2386         struct inode *inode = file_inode(file);
2387         struct address_space *mapping = inode->i_mapping;
2388         pgoff_t index;
2389         unsigned long offset;
2390         enum sgp_type sgp = SGP_READ;
2391         int error = 0;
2392         ssize_t retval = 0;
2393         loff_t *ppos = &iocb->ki_pos;
2394
2395         /*
2396          * Might this read be for a stacking filesystem?  Then when reading
2397          * holes of a sparse file, we actually need to allocate those pages,
2398          * and even mark them dirty, so it cannot exceed the max_blocks limit.
2399          */
2400         if (!iter_is_iovec(to))
2401                 sgp = SGP_CACHE;
2402
2403         index = *ppos >> PAGE_SHIFT;
2404         offset = *ppos & ~PAGE_MASK;
2405
2406         for (;;) {
2407                 struct page *page = NULL;
2408                 pgoff_t end_index;
2409                 unsigned long nr, ret;
2410                 loff_t i_size = i_size_read(inode);
2411
2412                 end_index = i_size >> PAGE_SHIFT;
2413                 if (index > end_index)
2414                         break;
2415                 if (index == end_index) {
2416                         nr = i_size & ~PAGE_MASK;
2417                         if (nr <= offset)
2418                                 break;
2419                 }
2420
2421                 error = shmem_getpage(inode, index, &page, sgp);
2422                 if (error) {
2423                         if (error == -EINVAL)
2424                                 error = 0;
2425                         break;
2426                 }
2427                 if (page) {
2428                         if (sgp == SGP_CACHE)
2429                                 set_page_dirty(page);
2430                         unlock_page(page);
2431                 }
2432
2433                 /*
2434                  * We must evaluate after, since reads (unlike writes)
2435                  * are called without i_mutex protection against truncate
2436                  */
2437                 nr = PAGE_SIZE;
2438                 i_size = i_size_read(inode);
2439                 end_index = i_size >> PAGE_SHIFT;
2440                 if (index == end_index) {
2441                         nr = i_size & ~PAGE_MASK;
2442                         if (nr <= offset) {
2443                                 if (page)
2444                                         put_page(page);
2445                                 break;
2446                         }
2447                 }
2448                 nr -= offset;
2449
2450                 if (page) {
2451                         /*
2452                          * If users can be writing to this page using arbitrary
2453                          * virtual addresses, take care about potential aliasing
2454                          * before reading the page on the kernel side.
2455                          */
2456                         if (mapping_writably_mapped(mapping))
2457                                 flush_dcache_page(page);
2458                         /*
2459                          * Mark the page accessed if we read the beginning.
2460                          */
2461                         if (!offset)
2462                                 mark_page_accessed(page);
2463                 } else {
2464                         page = ZERO_PAGE(0);
2465                         get_page(page);
2466                 }
2467
2468                 /*
2469                  * Ok, we have the page, and it's up-to-date, so
2470                  * now we can copy it to user space...
2471                  */
2472                 ret = copy_page_to_iter(page, offset, nr, to);
2473                 retval += ret;
2474                 offset += ret;
2475                 index += offset >> PAGE_SHIFT;
2476                 offset &= ~PAGE_MASK;
2477
2478                 put_page(page);
2479                 if (!iov_iter_count(to))
2480                         break;
2481                 if (ret < nr) {
2482                         error = -EFAULT;
2483                         break;
2484                 }
2485                 cond_resched();
2486         }
2487
2488         *ppos = ((loff_t) index << PAGE_SHIFT) + offset;
2489         file_accessed(file);
2490         return retval ? retval : error;
2491 }
2492
2493 /*
2494  * llseek SEEK_DATA or SEEK_HOLE through the radix_tree.
2495  */
2496 static pgoff_t shmem_seek_hole_data(struct address_space *mapping,
2497                                     pgoff_t index, pgoff_t end, int whence)
2498 {
2499         struct page *page;
2500         struct pagevec pvec;
2501         pgoff_t indices[PAGEVEC_SIZE];
2502         bool done = false;
2503         int i;
2504
2505         pagevec_init(&pvec, 0);
2506         pvec.nr = 1;            /* start small: we may be there already */
2507         while (!done) {
2508                 pvec.nr = find_get_entries(mapping, index,
2509                                         pvec.nr, pvec.pages, indices);
2510                 if (!pvec.nr) {
2511                         if (whence == SEEK_DATA)
2512                                 index = end;
2513                         break;
2514                 }
2515                 for (i = 0; i < pvec.nr; i++, index++) {
2516                         if (index < indices[i]) {
2517                                 if (whence == SEEK_HOLE) {
2518                                         done = true;
2519                                         break;
2520                                 }
2521                                 index = indices[i];
2522                         }
2523                         page = pvec.pages[i];
2524                         if (page && !radix_tree_exceptional_entry(page)) {
2525                                 if (!PageUptodate(page))
2526                                         page = NULL;
2527                         }
2528                         if (index >= end ||
2529                             (page && whence == SEEK_DATA) ||
2530                             (!page && whence == SEEK_HOLE)) {
2531                                 done = true;
2532                                 break;
2533                         }
2534                 }
2535                 pagevec_remove_exceptionals(&pvec);
2536                 pagevec_release(&pvec);
2537                 pvec.nr = PAGEVEC_SIZE;
2538                 cond_resched();
2539         }
2540         return index;
2541 }
2542
2543 static loff_t shmem_file_llseek(struct file *file, loff_t offset, int whence)
2544 {
2545         struct address_space *mapping = file->f_mapping;
2546         struct inode *inode = mapping->host;
2547         pgoff_t start, end;
2548         loff_t new_offset;
2549
2550         if (whence != SEEK_DATA && whence != SEEK_HOLE)
2551                 return generic_file_llseek_size(file, offset, whence,
2552                                         MAX_LFS_FILESIZE, i_size_read(inode));
2553         inode_lock(inode);
2554         /* We're holding i_mutex so we can access i_size directly */
2555
2556         if (offset < 0)
2557                 offset = -EINVAL;
2558         else if (offset >= inode->i_size)
2559                 offset = -ENXIO;
2560         else {
2561                 start = offset >> PAGE_SHIFT;
2562                 end = (inode->i_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2563                 new_offset = shmem_seek_hole_data(mapping, start, end, whence);
2564                 new_offset <<= PAGE_SHIFT;
2565                 if (new_offset > offset) {
2566                         if (new_offset < inode->i_size)
2567                                 offset = new_offset;
2568                         else if (whence == SEEK_DATA)
2569                                 offset = -ENXIO;
2570                         else
2571                                 offset = inode->i_size;
2572                 }
2573         }
2574
2575         if (offset >= 0)
2576                 offset = vfs_setpos(file, offset, MAX_LFS_FILESIZE);
2577         inode_unlock(inode);
2578         return offset;
2579 }
2580
2581 /*
2582  * We need a tag: a new tag would expand every radix_tree_node by 8 bytes,
2583  * so reuse a tag which we firmly believe is never set or cleared on shmem.
2584  */
2585 #define SHMEM_TAG_PINNED        PAGECACHE_TAG_TOWRITE
2586 #define LAST_SCAN               4       /* about 150ms max */
2587
2588 static void shmem_tag_pins(struct address_space *mapping)
2589 {
2590         struct radix_tree_iter iter;
2591         void **slot;
2592         pgoff_t start;
2593         struct page *page;
2594
2595         lru_add_drain();
2596         start = 0;
2597         rcu_read_lock();
2598
2599         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
2600                 page = radix_tree_deref_slot(slot);
2601                 if (!page || radix_tree_exception(page)) {
2602                         if (radix_tree_deref_retry(page)) {
2603                                 slot = radix_tree_iter_retry(&iter);
2604                                 continue;
2605                         }
2606                 } else if (page_count(page) - page_mapcount(page) > 1) {
2607                         spin_lock_irq(&mapping->tree_lock);
2608                         radix_tree_tag_set(&mapping->page_tree, iter.index,
2609                                            SHMEM_TAG_PINNED);
2610                         spin_unlock_irq(&mapping->tree_lock);
2611                 }
2612
2613                 if (need_resched()) {
2614                         slot = radix_tree_iter_resume(slot, &iter);
2615                         cond_resched_rcu();
2616                 }
2617         }
2618         rcu_read_unlock();
2619 }
2620
2621 /*
2622  * Setting SEAL_WRITE requires us to verify there's no pending writer. However,
2623  * via get_user_pages(), drivers might have some pending I/O without any active
2624  * user-space mappings (eg., direct-IO, AIO). Therefore, we look at all pages
2625  * and see whether it has an elevated ref-count. If so, we tag them and wait for
2626  * them to be dropped.
2627  * The caller must guarantee that no new user will acquire writable references
2628  * to those pages to avoid races.
2629  */
2630 static int shmem_wait_for_pins(struct address_space *mapping)
2631 {
2632         struct radix_tree_iter iter;
2633         void **slot;
2634         pgoff_t start;
2635         struct page *page;
2636         int error, scan;
2637
2638         shmem_tag_pins(mapping);
2639
2640         error = 0;
2641         for (scan = 0; scan <= LAST_SCAN; scan++) {
2642                 if (!radix_tree_tagged(&mapping->page_tree, SHMEM_TAG_PINNED))
2643                         break;
2644
2645                 if (!scan)
2646                         lru_add_drain_all();
2647                 else if (schedule_timeout_killable((HZ << scan) / 200))
2648                         scan = LAST_SCAN;
2649
2650                 start = 0;
2651                 rcu_read_lock();
2652                 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter,
2653                                            start, SHMEM_TAG_PINNED) {
2654
2655                         page = radix_tree_deref_slot(slot);
2656                         if (radix_tree_exception(page)) {
2657                                 if (radix_tree_deref_retry(page)) {
2658                                         slot = radix_tree_iter_retry(&iter);
2659                                         continue;
2660                                 }
2661
2662                                 page = NULL;
2663                         }
2664
2665                         if (page &&
2666                             page_count(page) - page_mapcount(page) != 1) {
2667                                 if (scan < LAST_SCAN)
2668                                         goto continue_resched;
2669
2670                                 /*
2671                                  * On the last scan, we clean up all those tags
2672                                  * we inserted; but make a note that we still
2673                                  * found pages pinned.
2674                                  */
2675                                 error = -EBUSY;
2676                         }
2677
2678                         spin_lock_irq(&mapping->tree_lock);
2679                         radix_tree_tag_clear(&mapping->page_tree,
2680                                              iter.index, SHMEM_TAG_PINNED);
2681                         spin_unlock_irq(&mapping->tree_lock);
2682 continue_resched:
2683                         if (need_resched()) {
2684                                 slot = radix_tree_iter_resume(slot, &iter);
2685                                 cond_resched_rcu();
2686                         }
2687                 }
2688                 rcu_read_unlock();
2689         }
2690
2691         return error;
2692 }
2693
2694 #define F_ALL_SEALS (F_SEAL_SEAL | \
2695                      F_SEAL_SHRINK | \
2696                      F_SEAL_GROW | \
2697                      F_SEAL_WRITE)
2698
2699 int shmem_add_seals(struct file *file, unsigned int seals)
2700 {
2701         struct inode *inode = file_inode(file);
2702         struct shmem_inode_info *info = SHMEM_I(inode);
2703         int error;
2704
2705         /*
2706          * SEALING
2707          * Sealing allows multiple parties to share a shmem-file but restrict
2708          * access to a specific subset of file operations. Seals can only be
2709          * added, but never removed. This way, mutually untrusted parties can
2710          * share common memory regions with a well-defined policy. A malicious
2711          * peer can thus never perform unwanted operations on a shared object.
2712          *
2713          * Seals are only supported on special shmem-files and always affect
2714          * the whole underlying inode. Once a seal is set, it may prevent some
2715          * kinds of access to the file. Currently, the following seals are
2716          * defined:
2717          *   SEAL_SEAL: Prevent further seals from being set on this file
2718          *   SEAL_SHRINK: Prevent the file from shrinking
2719          *   SEAL_GROW: Prevent the file from growing
2720          *   SEAL_WRITE: Prevent write access to the file
2721          *
2722          * As we don't require any trust relationship between two parties, we
2723          * must prevent seals from being removed. Therefore, sealing a file
2724          * only adds a given set of seals to the file, it never touches
2725          * existing seals. Furthermore, the "setting seals"-operation can be
2726          * sealed itself, which basically prevents any further seal from being
2727          * added.
2728          *
2729          * Semantics of sealing are only defined on volatile files. Only
2730          * anonymous shmem files support sealing. More importantly, seals are
2731          * never written to disk. Therefore, there's no plan to support it on
2732          * other file types.
2733          */
2734
2735         if (file->f_op != &shmem_file_operations)
2736                 return -EINVAL;
2737         if (!(file->f_mode & FMODE_WRITE))
2738                 return -EPERM;
2739         if (seals & ~(unsigned int)F_ALL_SEALS)
2740                 return -EINVAL;
2741
2742         inode_lock(inode);
2743
2744         if (info->seals & F_SEAL_SEAL) {
2745                 error = -EPERM;
2746                 goto unlock;
2747         }
2748
2749         if ((seals & F_SEAL_WRITE) && !(info->seals & F_SEAL_WRITE)) {
2750                 error = mapping_deny_writable(file->f_mapping);
2751                 if (error)
2752                         goto unlock;
2753
2754                 error = shmem_wait_for_pins(file->f_mapping);
2755                 if (error) {
2756                         mapping_allow_writable(file->f_mapping);
2757                         goto unlock;
2758                 }
2759         }
2760
2761         info->seals |= seals;
2762         error = 0;
2763
2764 unlock:
2765         inode_unlock(inode);
2766         return error;
2767 }
2768 EXPORT_SYMBOL_GPL(shmem_add_seals);
2769
2770 int shmem_get_seals(struct file *file)
2771 {
2772         if (file->f_op != &shmem_file_operations)
2773                 return -EINVAL;
2774
2775         return SHMEM_I(file_inode(file))->seals;
2776 }
2777 EXPORT_SYMBOL_GPL(shmem_get_seals);
2778
2779 long shmem_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
2780 {
2781         long error;
2782
2783         switch (cmd) {
2784         case F_ADD_SEALS:
2785                 /* disallow upper 32bit */
2786                 if (arg > UINT_MAX)
2787                         return -EINVAL;
2788
2789                 error = shmem_add_seals(file, arg);
2790                 break;
2791         case F_GET_SEALS:
2792                 error = shmem_get_seals(file);
2793                 break;
2794         default:
2795                 error = -EINVAL;
2796                 break;
2797         }
2798
2799         return error;
2800 }
2801
2802 static long shmem_fallocate(struct file *file, int mode, loff_t offset,
2803                                                          loff_t len)
2804 {
2805         struct inode *inode = file_inode(file);
2806         struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
2807         struct shmem_inode_info *info = SHMEM_I(inode);
2808         struct shmem_falloc shmem_falloc;
2809         pgoff_t start, index, end;
2810         int error;
2811
2812         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2813                 return -EOPNOTSUPP;
2814
2815         inode_lock(inode);
2816
2817         if (mode & FALLOC_FL_PUNCH_HOLE) {
2818                 struct address_space *mapping = file->f_mapping;
2819                 loff_t unmap_start = round_up(offset, PAGE_SIZE);
2820                 loff_t unmap_end = round_down(offset + len, PAGE_SIZE) - 1;
2821                 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(shmem_falloc_waitq);
2822
2823                 /* protected by i_mutex */
2824                 if (info->seals & F_SEAL_WRITE) {
2825                         error = -EPERM;
2826                         goto out;
2827                 }
2828
2829                 shmem_falloc.waitq = &shmem_falloc_waitq;
2830                 shmem_falloc.start = unmap_start >> PAGE_SHIFT;
2831                 shmem_falloc.next = (unmap_end + 1) >> PAGE_SHIFT;
2832                 spin_lock(&inode->i_lock);
2833                 inode->i_private = &shmem_falloc;
2834                 spin_unlock(&inode->i_lock);
2835
2836                 if ((u64)unmap_end > (u64)unmap_start)
2837                         unmap_mapping_range(mapping, unmap_start,
2838                                             1 + unmap_end - unmap_start, 0);
2839                 shmem_truncate_range(inode, offset, offset + len - 1);
2840                 /* No need to unmap again: hole-punching leaves COWed pages */
2841
2842                 spin_lock(&inode->i_lock);
2843                 inode->i_private = NULL;
2844                 wake_up_all(&shmem_falloc_waitq);
2845                 WARN_ON_ONCE(!list_empty(&shmem_falloc_waitq.head));
2846                 spin_unlock(&inode->i_lock);
2847                 error = 0;
2848                 goto out;
2849         }
2850
2851         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
2852         error = inode_newsize_ok(inode, offset + len);
2853         if (error)
2854                 goto out;
2855
2856         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
2857                 error = -EPERM;
2858                 goto out;
2859         }
2860
2861         start = offset >> PAGE_SHIFT;
2862         end = (offset + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2863         /* Try to avoid a swapstorm if len is impossible to satisfy */
2864         if (sbinfo->max_blocks && end - start > sbinfo->max_blocks) {
2865                 error = -ENOSPC;
2866                 goto out;
2867         }
2868
2869         shmem_falloc.waitq = NULL;
2870         shmem_falloc.start = start;
2871         shmem_falloc.next  = start;
2872         shmem_falloc.nr_falloced = 0;
2873         shmem_falloc.nr_unswapped = 0;
2874         spin_lock(&inode->i_lock);
2875         inode->i_private = &shmem_falloc;
2876         spin_unlock(&inode->i_lock);
2877
2878         for (index = start; index < end; index++) {
2879                 struct page *page;
2880
2881                 /*
2882                  * Good, the fallocate(2) manpage permits EINTR: we may have
2883                  * been interrupted because we are using up too much memory.
2884                  */
2885                 if (signal_pending(current))
2886                         error = -EINTR;
2887                 else if (shmem_falloc.nr_unswapped > shmem_falloc.nr_falloced)
2888                         error = -ENOMEM;
2889                 else
2890                         error = shmem_getpage(inode, index, &page, SGP_FALLOC);
2891                 if (error) {
2892                         /* Remove the !PageUptodate pages we added */
2893                         if (index > start) {
2894                                 shmem_undo_range(inode,
2895                                     (loff_t)start << PAGE_SHIFT,
2896                                     ((loff_t)index << PAGE_SHIFT) - 1, true);
2897                         }
2898                         goto undone;
2899                 }
2900
2901                 /*
2902                  * Inform shmem_writepage() how far we have reached.
2903                  * No need for lock or barrier: we have the page lock.
2904                  */
2905                 shmem_falloc.next++;
2906                 if (!PageUptodate(page))
2907                         shmem_falloc.nr_falloced++;
2908
2909                 /*
2910                  * If !PageUptodate, leave it that way so that freeable pages
2911                  * can be recognized if we need to rollback on error later.
2912                  * But set_page_dirty so that memory pressure will swap rather
2913                  * than free the pages we are allocating (and SGP_CACHE pages
2914                  * might still be clean: we now need to mark those dirty too).
2915                  */
2916                 set_page_dirty(page);
2917                 unlock_page(page);
2918                 put_page(page);
2919                 cond_resched();
2920         }
2921
2922         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
2923                 i_size_write(inode, offset + len);
2924         inode->i_ctime = current_time(inode);
2925 undone:
2926         spin_lock(&inode->i_lock);
2927         inode->i_private = NULL;
2928         spin_unlock(&inode->i_lock);
2929 out:
2930         inode_unlock(inode);
2931         return error;
2932 }
2933
2934 static int shmem_statfs(struct dentry *dentry, struct kstatfs *buf)
2935 {
2936         struct shmem_sb_info *sbinfo = SHMEM_SB(dentry->d_sb);
2937
2938         buf->f_type = TMPFS_MAGIC;
2939         buf->f_bsize = PAGE_SIZE;
2940         buf->f_namelen = NAME_MAX;
2941         if (sbinfo->max_blocks) {
2942                 buf->f_blocks = sbinfo->max_blocks;
2943                 buf->f_bavail =
2944                 buf->f_bfree  = sbinfo->max_blocks -
2945                                 percpu_counter_sum(&sbinfo->used_blocks);
2946         }
2947         if (sbinfo->max_inodes) {
2948                 buf->f_files = sbinfo->max_inodes;
2949                 buf->f_ffree = sbinfo->free_inodes;
2950         }
2951         /* else leave those fields 0 like simple_statfs */
2952         return 0;
2953 }
2954
2955 /*
2956  * File creation. Allocate an inode, and we're done..
2957  */
2958 static int
2959 shmem_mknod(struct inode *dir, struct dentry *dentry, umode_t mode, dev_t dev)
2960 {
2961         struct inode *inode;
2962         int error = -ENOSPC;
2963
2964         inode = shmem_get_inode(dir->i_sb, dir, mode, dev, VM_NORESERVE);
2965         if (inode) {
2966                 error = simple_acl_create(dir, inode);
2967                 if (error)
2968                         goto out_iput;
2969                 error = security_inode_init_security(inode, dir,
2970                                                      &dentry->d_name,
2971                                                      shmem_initxattrs, NULL);
2972                 if (error && error != -EOPNOTSUPP)
2973                         goto out_iput;
2974
2975                 error = 0;
2976                 dir->i_size += BOGO_DIRENT_SIZE;
2977                 dir->i_ctime = dir->i_mtime = current_time(dir);
2978                 d_instantiate(dentry, inode);
2979                 dget(dentry); /* Extra count - pin the dentry in core */
2980         }
2981         return error;
2982 out_iput:
2983         iput(inode);
2984         return error;
2985 }
2986
2987 static int
2988 shmem_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
2989 {
2990         struct inode *inode;
2991         int error = -ENOSPC;
2992
2993         inode = shmem_get_inode(dir->i_sb, dir, mode, 0, VM_NORESERVE);
2994         if (inode) {
2995                 error = security_inode_init_security(inode, dir,
2996                                                      NULL,
2997                                                      shmem_initxattrs, NULL);
2998                 if (error && error != -EOPNOTSUPP)
2999                         goto out_iput;
3000                 error = simple_acl_create(dir, inode);
3001                 if (error)
3002                         goto out_iput;
3003                 d_tmpfile(dentry, inode);
3004         }
3005         return error;
3006 out_iput:
3007         iput(inode);
3008         return error;
3009 }
3010
3011 static int shmem_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
3012 {
3013         int error;
3014
3015         if ((error = shmem_mknod(dir, dentry, mode | S_IFDIR, 0)))
3016                 return error;
3017         inc_nlink(dir);
3018         return 0;
3019 }
3020
3021 static int shmem_create(struct inode *dir, struct dentry *dentry, umode_t mode,
3022                 bool excl)
3023 {
3024         return shmem_mknod(dir, dentry, mode | S_IFREG, 0);
3025 }
3026
3027 /*
3028  * Link a file..
3029  */
3030 static int shmem_link(struct dentry *old_dentry, struct inode *dir, struct dentry *dentry)
3031 {
3032         struct inode *inode = d_inode(old_dentry);
3033         int ret;
3034
3035         /*
3036          * No ordinary (disk based) filesystem counts links as inodes;
3037          * but each new link needs a new dentry, pinning lowmem, and
3038          * tmpfs dentries cannot be pruned until they are unlinked.
3039          */
3040         ret = shmem_reserve_inode(inode->i_sb);
3041         if (ret)
3042                 goto out;
3043
3044         dir->i_size += BOGO_DIRENT_SIZE;
3045         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3046         inc_nlink(inode);
3047         ihold(inode);   /* New dentry reference */
3048         dget(dentry);           /* Extra pinning count for the created dentry */
3049         d_instantiate(dentry, inode);
3050 out:
3051         return ret;
3052 }
3053
3054 static int shmem_unlink(struct inode *dir, struct dentry *dentry)
3055 {
3056         struct inode *inode = d_inode(dentry);
3057
3058         if (inode->i_nlink > 1 && !S_ISDIR(inode->i_mode))
3059                 shmem_free_inode(inode->i_sb);
3060
3061         dir->i_size -= BOGO_DIRENT_SIZE;
3062         inode->i_ctime = dir->i_ctime = dir->i_mtime = current_time(inode);
3063         drop_nlink(inode);
3064         dput(dentry);   /* Undo the count from "create" - this does all the work */
3065         return 0;
3066 }
3067
3068 static int shmem_rmdir(struct inode *dir, struct dentry *dentry)
3069 {
3070         if (!simple_empty(dentry))
3071                 return -ENOTEMPTY;
3072
3073         drop_nlink(d_inode(dentry));
3074         drop_nlink(dir);
3075         return shmem_unlink(dir, dentry);
3076 }
3077
3078 static int shmem_exchange(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry)
3079 {
3080         bool old_is_dir = d_is_dir(old_dentry);
3081         bool new_is_dir = d_is_dir(new_dentry);
3082
3083         if (old_dir != new_dir && old_is_dir != new_is_dir) {
3084                 if (old_is_dir) {
3085                         drop_nlink(old_dir);
3086                         inc_nlink(new_dir);
3087                 } else {
3088                         drop_nlink(new_dir);
3089                         inc_nlink(old_dir);
3090                 }
3091         }
3092         old_dir->i_ctime = old_dir->i_mtime =
3093         new_dir->i_ctime = new_dir->i_mtime =
3094         d_inode(old_dentry)->i_ctime =
3095         d_inode(new_dentry)->i_ctime = current_time(old_dir);
3096
3097         return 0;
3098 }
3099
3100 static int shmem_whiteout(struct inode *old_dir, struct dentry *old_dentry)
3101 {
3102         struct dentry *whiteout;
3103         int error;
3104
3105         whiteout = d_alloc(old_dentry->d_parent, &old_dentry->d_name);
3106         if (!whiteout)
3107                 return -ENOMEM;
3108
3109         error = shmem_mknod(old_dir, whiteout,
3110                             S_IFCHR | WHITEOUT_MODE, WHITEOUT_DEV);
3111         dput(whiteout);
3112         if (error)
3113                 return error;
3114
3115         /*
3116          * Cheat and hash the whiteout while the old dentry is still in
3117          * place, instead of playing games with FS_RENAME_DOES_D_MOVE.
3118          *
3119          * d_lookup() will consistently find one of them at this point,
3120          * not sure which one, but that isn't even important.
3121          */
3122         d_rehash(whiteout);
3123         return 0;
3124 }
3125
3126 /*
3127  * The VFS layer already does all the dentry stuff for rename,
3128  * we just have to decrement the usage count for the target if
3129  * it exists so that the VFS layer correctly free's it when it
3130  * gets overwritten.
3131  */
3132 static int shmem_rename2(struct inode *old_dir, struct dentry *old_dentry, struct inode *new_dir, struct dentry *new_dentry, unsigned int flags)
3133 {
3134         struct inode *inode = d_inode(old_dentry);
3135         int they_are_dirs = S_ISDIR(inode->i_mode);
3136
3137         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
3138                 return -EINVAL;
3139
3140         if (flags & RENAME_EXCHANGE)
3141                 return shmem_exchange(old_dir, old_dentry, new_dir, new_dentry);
3142
3143         if (!simple_empty(new_dentry))
3144                 return -ENOTEMPTY;
3145
3146         if (flags & RENAME_WHITEOUT) {
3147                 int error;
3148
3149                 error = shmem_whiteout(old_dir, old_dentry);
3150                 if (error)
3151                         return error;
3152         }
3153
3154         if (d_really_is_positive(new_dentry)) {
3155                 (void) shmem_unlink(new_dir, new_dentry);
3156                 if (they_are_dirs) {
3157                         drop_nlink(d_inode(new_dentry));
3158                         drop_nlink(old_dir);
3159                 }
3160         } else if (they_are_dirs) {
3161                 drop_nlink(old_dir);
3162                 inc_nlink(new_dir);
3163         }
3164
3165         old_dir->i_size -= BOGO_DIRENT_SIZE;
3166         new_dir->i_size += BOGO_DIRENT_SIZE;
3167         old_dir->i_ctime = old_dir->i_mtime =
3168         new_dir->i_ctime = new_dir->i_mtime =
3169         inode->i_ctime = current_time(old_dir);
3170         return 0;
3171 }
3172
3173 static int shmem_symlink(struct inode *dir, struct dentry *dentry, const char *symname)
3174 {
3175         int error;
3176         int len;
3177         struct inode *inode;
3178         struct page *page;
3179         struct shmem_inode_info *info;
3180
3181         len = strlen(symname) + 1;
3182         if (len > PAGE_SIZE)
3183                 return -ENAMETOOLONG;
3184
3185         inode = shmem_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0, VM_NORESERVE);
3186         if (!inode)
3187                 return -ENOSPC;
3188
3189         error = security_inode_init_security(inode, dir, &dentry->d_name,
3190                                              shmem_initxattrs, NULL);
3191         if (error) {
3192                 if (error != -EOPNOTSUPP) {
3193                         iput(inode);
3194                         return error;
3195                 }
3196                 error = 0;
3197         }
3198
3199         info = SHMEM_I(inode);
3200         inode->i_size = len-1;
3201         if (len <= SHORT_SYMLINK_LEN) {
3202                 inode->i_link = kmemdup(symname, len, GFP_KERNEL);
3203                 if (!inode->i_link) {
3204                         iput(inode);
3205                         return -ENOMEM;
3206                 }
3207                 inode->i_op = &shmem_short_symlink_operations;
3208         } else {
3209                 inode_nohighmem(inode);
3210                 error = shmem_getpage(inode, 0, &page, SGP_WRITE);
3211                 if (error) {
3212                         iput(inode);
3213                         return error;
3214                 }
3215                 inode->i_mapping->a_ops = &shmem_aops;
3216                 inode->i_op = &shmem_symlink_inode_operations;
3217                 memcpy(page_address(page), symname, len);
3218                 SetPageUptodate(page);
3219                 set_page_dirty(page);
3220                 unlock_page(page);
3221                 put_page(page);
3222         }
3223         dir->i_size += BOGO_DIRENT_SIZE;
3224         dir->i_ctime = dir->i_mtime = current_time(dir);
3225         d_instantiate(dentry, inode);
3226         dget(dentry);
3227         return 0;
3228 }
3229
3230 static void shmem_put_link(void *arg)
3231 {
3232         mark_page_accessed(arg);
3233         put_page(arg);
3234 }
3235
3236 static const char *shmem_get_link(struct dentry *dentry,
3237                                   struct inode *inode,
3238                                   struct delayed_call *done)
3239 {
3240         struct page *page = NULL;
3241         int error;
3242         if (!dentry) {
3243                 page = find_get_page(inode->i_mapping, 0);
3244                 if (!page)
3245                         return ERR_PTR(-ECHILD);
3246                 if (!PageUptodate(page)) {
3247                         put_page(page);
3248                         return ERR_PTR(-ECHILD);
3249                 }
3250         } else {
3251                 error = shmem_getpage(inode, 0, &page, SGP_READ);
3252                 if (error)
3253                         return ERR_PTR(error);
3254                 unlock_page(page);
3255         }
3256         set_delayed_call(done, shmem_put_link, page);
3257         return page_address(page);
3258 }
3259
3260 #ifdef CONFIG_TMPFS_XATTR
3261 /*
3262  * Superblocks without xattr inode operations may get some security.* xattr
3263  * support from the LSM "for free". As soon as we have any other xattrs
3264  * like ACLs, we also need to implement the security.* handlers at
3265  * filesystem level, though.
3266  */
3267
3268 /*
3269  * Callback for security_inode_init_security() for acquiring xattrs.
3270  */
3271 static int shmem_initxattrs(struct inode *inode,
3272                             const struct xattr *xattr_array,
3273                             void *fs_info)
3274 {
3275         struct shmem_inode_info *info = SHMEM_I(inode);
3276         const struct xattr *xattr;
3277         struct simple_xattr *new_xattr;
3278         size_t len;
3279
3280         for (xattr = xattr_array; xattr->name != NULL; xattr++) {
3281                 new_xattr = simple_xattr_alloc(xattr->value, xattr->value_len);
3282                 if (!new_xattr)
3283                         return -ENOMEM;
3284
3285                 len = strlen(xattr->name) + 1;
3286                 new_xattr->name = kmalloc(XATTR_SECURITY_PREFIX_LEN + len,
3287                                           GFP_KERNEL);
3288                 if (!new_xattr->name) {
3289                         kfree(new_xattr);
3290                         return -ENOMEM;
3291                 }
3292
3293                 memcpy(new_xattr->name, XATTR_SECURITY_PREFIX,
3294                        XATTR_SECURITY_PREFIX_LEN);
3295                 memcpy(new_xattr->name + XATTR_SECURITY_PREFIX_LEN,
3296                        xattr->name, len);
3297
3298                 simple_xattr_list_add(&info->xattrs, new_xattr);
3299         }
3300
3301         return 0;
3302 }
3303
3304 static int shmem_xattr_handler_get(const struct xattr_handler *handler,
3305                                    struct dentry *unused, struct inode *inode,
3306                                    const char *name, void *buffer, size_t size)
3307 {
3308         struct shmem_inode_info *info = SHMEM_I(inode);
3309
3310         name = xattr_full_name(handler, name);
3311         return simple_xattr_get(&info->xattrs, name, buffer, size);
3312 }
3313
3314 static int shmem_xattr_handler_set(const struct xattr_handler *handler,
3315                                    struct dentry *unused, struct inode *inode,
3316                                    const char *name, const void *value,
3317                                    size_t size, int flags)
3318 {
3319         struct shmem_inode_info *info = SHMEM_I(inode);
3320
3321         name = xattr_full_name(handler, name);
3322         return simple_xattr_set(&info->xattrs, name, value, size, flags);
3323 }
3324
3325 static const struct xattr_handler shmem_security_xattr_handler = {
3326         .prefix = XATTR_SECURITY_PREFIX,
3327         .get = shmem_xattr_handler_get,
3328         .set = shmem_xattr_handler_set,
3329 };
3330
3331 static const struct xattr_handler shmem_trusted_xattr_handler = {
3332         .prefix = XATTR_TRUSTED_PREFIX,
3333         .get = shmem_xattr_handler_get,
3334         .set = shmem_xattr_handler_set,
3335 };
3336
3337 static const struct xattr_handler *shmem_xattr_handlers[] = {
3338 #ifdef CONFIG_TMPFS_POSIX_ACL
3339         &posix_acl_access_xattr_handler,
3340         &posix_acl_default_xattr_handler,
3341 #endif
3342         &shmem_security_xattr_handler,
3343         &shmem_trusted_xattr_handler,
3344         NULL
3345 };
3346
3347 static ssize_t shmem_listxattr(struct dentry *dentry, char *buffer, size_t size)
3348 {
3349         struct shmem_inode_info *info = SHMEM_I(d_inode(dentry));
3350         return simple_xattr_list(d_inode(dentry), &info->xattrs, buffer, size);
3351 }
3352 #endif /* CONFIG_TMPFS_XATTR */
3353
3354 static const struct inode_operations shmem_short_symlink_operations = {
3355         .get_link       = simple_get_link,
3356 #ifdef CONFIG_TMPFS_XATTR
3357         .listxattr      = shmem_listxattr,
3358 #endif
3359 };
3360
3361 static const struct inode_operations shmem_symlink_inode_operations = {
3362         .get_link       = shmem_get_link,
3363 #ifdef CONFIG_TMPFS_XATTR
3364         .listxattr      = shmem_listxattr,
3365 #endif
3366 };
3367
3368 static struct dentry *shmem_get_parent(struct dentry *child)
3369 {
3370         return ERR_PTR(-ESTALE);
3371 }
3372
3373 static int shmem_match(struct inode *ino, void *vfh)
3374 {
3375         __u32 *fh = vfh;
3376         __u64 inum = fh[2];
3377         inum = (inum << 32) | fh[1];
3378         return ino->i_ino == inum && fh[0] == ino->i_generation;
3379 }
3380
3381 static struct dentry *shmem_fh_to_dentry(struct super_block *sb,
3382                 struct fid *fid, int fh_len, int fh_type)
3383 {
3384         struct inode *inode;
3385         struct dentry *dentry = NULL;
3386         u64 inum;
3387
3388         if (fh_len < 3)
3389                 return NULL;
3390
3391         inum = fid->raw[2];
3392         inum = (inum << 32) | fid->raw[1];
3393
3394         inode = ilookup5(sb, (unsigned long)(inum + fid->raw[0]),
3395                         shmem_match, fid->raw);
3396         if (inode) {
3397                 dentry = d_find_alias(inode);
3398                 iput(inode);
3399         }
3400
3401         return dentry;
3402 }
3403
3404 static int shmem_encode_fh(struct inode *inode, __u32 *fh, int *len,
3405                                 struct inode *parent)
3406 {
3407         if (*len < 3) {
3408                 *len = 3;
3409                 return FILEID_INVALID;
3410         }
3411
3412         if (inode_unhashed(inode)) {
3413                 /* Unfortunately insert_inode_hash is not idempotent,
3414                  * so as we hash inodes here rather than at creation
3415                  * time, we need a lock to ensure we only try
3416                  * to do it once
3417                  */
3418                 static DEFINE_SPINLOCK(lock);
3419                 spin_lock(&lock);
3420                 if (inode_unhashed(inode))
3421                         __insert_inode_hash(inode,
3422                                             inode->i_ino + inode->i_generation);
3423                 spin_unlock(&lock);
3424         }
3425
3426         fh[0] = inode->i_generation;
3427         fh[1] = inode->i_ino;
3428         fh[2] = ((__u64)inode->i_ino) >> 32;
3429
3430         *len = 3;
3431         return 1;
3432 }
3433
3434 static const struct export_operations shmem_export_ops = {
3435         .get_parent     = shmem_get_parent,
3436         .encode_fh      = shmem_encode_fh,
3437         .fh_to_dentry   = shmem_fh_to_dentry,
3438 };
3439
3440 static int shmem_parse_options(char *options, struct shmem_sb_info *sbinfo,
3441                                bool remount)
3442 {
3443         char *this_char, *value, *rest;
3444         struct mempolicy *mpol = NULL;
3445         uid_t uid;
3446         gid_t gid;
3447
3448         while (options != NULL) {
3449                 this_char = options;
3450                 for (;;) {
3451                         /*
3452                          * NUL-terminate this option: unfortunately,
3453                          * mount options form a comma-separated list,
3454                          * but mpol's nodelist may also contain commas.
3455                          */
3456                         options = strchr(options, ',');
3457                         if (options == NULL)
3458                                 break;
3459                         options++;
3460                         if (!isdigit(*options)) {
3461                                 options[-1] = '\0';
3462                                 break;
3463                         }
3464                 }
3465                 if (!*this_char)
3466                         continue;
3467                 if ((value = strchr(this_char,'=')) != NULL) {
3468                         *value++ = 0;
3469                 } else {
3470                         pr_err("tmpfs: No value for mount option '%s'\n",
3471                                this_char);
3472                         goto error;
3473                 }
3474
3475                 if (!strcmp(this_char,"size")) {
3476                         unsigned long long size;
3477                         size = memparse(value,&rest);
3478                         if (*rest == '%') {
3479                                 size <<= PAGE_SHIFT;
3480                                 size *= totalram_pages;
3481                                 do_div(size, 100);
3482                                 rest++;
3483                         }
3484                         if (*rest)
3485                                 goto bad_val;
3486                         sbinfo->max_blocks =
3487                                 DIV_ROUND_UP(size, PAGE_SIZE);
3488                 } else if (!strcmp(this_char,"nr_blocks")) {
3489                         sbinfo->max_blocks = memparse(value, &rest);
3490                         if (*rest)
3491                                 goto bad_val;
3492                 } else if (!strcmp(this_char,"nr_inodes")) {
3493                         sbinfo->max_inodes = memparse(value, &rest);
3494                         if (*rest)
3495                                 goto bad_val;
3496                 } else if (!strcmp(this_char,"mode")) {
3497                         if (remount)
3498                                 continue;
3499                         sbinfo->mode = simple_strtoul(value, &rest, 8) & 07777;
3500                         if (*rest)
3501                                 goto bad_val;
3502                 } else if (!strcmp(this_char,"uid")) {
3503                         if (remount)
3504                                 continue;
3505                         uid = simple_strtoul(value, &rest, 0);
3506                         if (*rest)
3507                                 goto bad_val;
3508                         sbinfo->uid = make_kuid(current_user_ns(), uid);
3509                         if (!uid_valid(sbinfo->uid))
3510                                 goto bad_val;
3511                 } else if (!strcmp(this_char,"gid")) {
3512                         if (remount)
3513                                 continue;
3514                         gid = simple_strtoul(value, &rest, 0);
3515                         if (*rest)
3516                                 goto bad_val;
3517                         sbinfo->gid = make_kgid(current_user_ns(), gid);
3518                         if (!gid_valid(sbinfo->gid))
3519                                 goto bad_val;
3520 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3521                 } else if (!strcmp(this_char, "huge")) {
3522                         int huge;
3523                         huge = shmem_parse_huge(value);
3524                         if (huge < 0)
3525                                 goto bad_val;
3526                         if (!has_transparent_hugepage() &&
3527                                         huge != SHMEM_HUGE_NEVER)
3528                                 goto bad_val;
3529                         sbinfo->huge = huge;
3530 #endif
3531 #ifdef CONFIG_NUMA
3532                 } else if (!strcmp(this_char,"mpol")) {
3533                         mpol_put(mpol);
3534                         mpol = NULL;
3535                         if (mpol_parse_str(value, &mpol))
3536                                 goto bad_val;
3537 #endif
3538                 } else {
3539                         pr_err("tmpfs: Bad mount option %s\n", this_char);
3540                         goto error;
3541                 }
3542         }
3543         sbinfo->mpol = mpol;
3544         return 0;
3545
3546 bad_val:
3547         pr_err("tmpfs: Bad value '%s' for mount option '%s'\n",
3548                value, this_char);
3549 error:
3550         mpol_put(mpol);
3551         return 1;
3552
3553 }
3554
3555 static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
3556 {
3557         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3558         struct shmem_sb_info config = *sbinfo;
3559         unsigned long inodes;
3560         int error = -EINVAL;
3561
3562         config.mpol = NULL;
3563         if (shmem_parse_options(data, &config, true))
3564                 return error;
3565
3566         spin_lock(&sbinfo->stat_lock);
3567         inodes = sbinfo->max_inodes - sbinfo->free_inodes;
3568         if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
3569                 goto out;
3570         if (config.max_inodes < inodes)
3571                 goto out;
3572         /*
3573          * Those tests disallow limited->unlimited while any are in use;
3574          * but we must separately disallow unlimited->limited, because
3575          * in that case we have no record of how much is already in use.
3576          */
3577         if (config.max_blocks && !sbinfo->max_blocks)
3578                 goto out;
3579         if (config.max_inodes && !sbinfo->max_inodes)
3580                 goto out;
3581
3582         error = 0;
3583         sbinfo->huge = config.huge;
3584         sbinfo->max_blocks  = config.max_blocks;
3585         sbinfo->max_inodes  = config.max_inodes;
3586         sbinfo->free_inodes = config.max_inodes - inodes;
3587
3588         /*
3589          * Preserve previous mempolicy unless mpol remount option was specified.
3590          */
3591         if (config.mpol) {
3592                 mpol_put(sbinfo->mpol);
3593                 sbinfo->mpol = config.mpol;     /* transfers initial ref */
3594         }
3595 out:
3596         spin_unlock(&sbinfo->stat_lock);
3597         return error;
3598 }
3599
3600 static int shmem_show_options(struct seq_file *seq, struct dentry *root)
3601 {
3602         struct shmem_sb_info *sbinfo = SHMEM_SB(root->d_sb);
3603
3604         if (sbinfo->max_blocks != shmem_default_max_blocks())
3605                 seq_printf(seq, ",size=%luk",
3606                         sbinfo->max_blocks << (PAGE_SHIFT - 10));
3607         if (sbinfo->max_inodes != shmem_default_max_inodes())
3608                 seq_printf(seq, ",nr_inodes=%lu", sbinfo->max_inodes);
3609         if (sbinfo->mode != (S_IRWXUGO | S_ISVTX))
3610                 seq_printf(seq, ",mode=%03ho", sbinfo->mode);
3611         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
3612                 seq_printf(seq, ",uid=%u",
3613                                 from_kuid_munged(&init_user_ns, sbinfo->uid));
3614         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
3615                 seq_printf(seq, ",gid=%u",
3616                                 from_kgid_munged(&init_user_ns, sbinfo->gid));
3617 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3618         /* Rightly or wrongly, show huge mount option unmasked by shmem_huge */
3619         if (sbinfo->huge)
3620                 seq_printf(seq, ",huge=%s", shmem_format_huge(sbinfo->huge));
3621 #endif
3622         shmem_show_mpol(seq, sbinfo->mpol);
3623         return 0;
3624 }
3625
3626 #define MFD_NAME_PREFIX "memfd:"
3627 #define MFD_NAME_PREFIX_LEN (sizeof(MFD_NAME_PREFIX) - 1)
3628 #define MFD_NAME_MAX_LEN (NAME_MAX - MFD_NAME_PREFIX_LEN)
3629
3630 #define MFD_ALL_FLAGS (MFD_CLOEXEC | MFD_ALLOW_SEALING)
3631
3632 SYSCALL_DEFINE2(memfd_create,
3633                 const char __user *, uname,
3634                 unsigned int, flags)
3635 {
3636         struct shmem_inode_info *info;
3637         struct file *file;
3638         int fd, error;
3639         char *name;
3640         long len;
3641
3642         if (flags & ~(unsigned int)MFD_ALL_FLAGS)
3643                 return -EINVAL;
3644
3645         /* length includes terminating zero */
3646         len = strnlen_user(uname, MFD_NAME_MAX_LEN + 1);
3647         if (len <= 0)
3648                 return -EFAULT;
3649         if (len > MFD_NAME_MAX_LEN + 1)
3650                 return -EINVAL;
3651
3652         name = kmalloc(len + MFD_NAME_PREFIX_LEN, GFP_TEMPORARY);
3653         if (!name)
3654                 return -ENOMEM;
3655
3656         strcpy(name, MFD_NAME_PREFIX);
3657         if (copy_from_user(&name[MFD_NAME_PREFIX_LEN], uname, len)) {
3658                 error = -EFAULT;
3659                 goto err_name;
3660         }
3661
3662         /* terminating-zero may have changed after strnlen_user() returned */
3663         if (name[len + MFD_NAME_PREFIX_LEN - 1]) {
3664                 error = -EFAULT;
3665                 goto err_name;
3666         }
3667
3668         fd = get_unused_fd_flags((flags & MFD_CLOEXEC) ? O_CLOEXEC : 0);
3669         if (fd < 0) {
3670                 error = fd;
3671                 goto err_name;
3672         }
3673
3674         file = shmem_file_setup(name, 0, VM_NORESERVE);
3675         if (IS_ERR(file)) {
3676                 error = PTR_ERR(file);
3677                 goto err_fd;
3678         }
3679         info = SHMEM_I(file_inode(file));
3680         file->f_mode |= FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE;
3681         file->f_flags |= O_RDWR | O_LARGEFILE;
3682         if (flags & MFD_ALLOW_SEALING)
3683                 info->seals &= ~F_SEAL_SEAL;
3684
3685         fd_install(fd, file);
3686         kfree(name);
3687         return fd;
3688
3689 err_fd:
3690         put_unused_fd(fd);
3691 err_name:
3692         kfree(name);
3693         return error;
3694 }
3695
3696 #endif /* CONFIG_TMPFS */
3697
3698 static void shmem_put_super(struct super_block *sb)
3699 {
3700         struct shmem_sb_info *sbinfo = SHMEM_SB(sb);
3701
3702         percpu_counter_destroy(&sbinfo->used_blocks);
3703         mpol_put(sbinfo->mpol);
3704         kfree(sbinfo);
3705         sb->s_fs_info = NULL;
3706 }
3707
3708 int shmem_fill_super(struct super_block *sb, void *data, int silent)
3709 {
3710         struct inode *inode;
3711         struct shmem_sb_info *sbinfo;
3712         int err = -ENOMEM;
3713
3714         /* Round up to L1_CACHE_BYTES to resist false sharing */
3715         sbinfo = kzalloc(max((int)sizeof(struct shmem_sb_info),
3716                                 L1_CACHE_BYTES), GFP_KERNEL);
3717         if (!sbinfo)
3718                 return -ENOMEM;
3719
3720         sbinfo->mode = S_IRWXUGO | S_ISVTX;
3721         sbinfo->uid = current_fsuid();
3722         sbinfo->gid = current_fsgid();
3723         sb->s_fs_info = sbinfo;
3724
3725 #ifdef CONFIG_TMPFS
3726         /*
3727          * Per default we only allow half of the physical ram per
3728          * tmpfs instance, limiting inodes to one per page of lowmem;
3729          * but the internal instance is left unlimited.
3730          */
3731         if (!(sb->s_flags & MS_KERNMOUNT)) {
3732                 sbinfo->max_blocks = shmem_default_max_blocks();
3733                 sbinfo->max_inodes = shmem_default_max_inodes();
3734                 if (shmem_parse_options(data, sbinfo, false)) {
3735                         err = -EINVAL;
3736                         goto failed;
3737                 }
3738         } else {
3739                 sb->s_flags |= MS_NOUSER;
3740         }
3741         sb->s_export_op = &shmem_export_ops;
3742         sb->s_flags |= MS_NOSEC;
3743 #else
3744         sb->s_flags |= MS_NOUSER;
3745 #endif
3746
3747         spin_lock_init(&sbinfo->stat_lock);
3748         if (percpu_counter_init(&sbinfo->used_blocks, 0, GFP_KERNEL))
3749                 goto failed;
3750         sbinfo->free_inodes = sbinfo->max_inodes;
3751         spin_lock_init(&sbinfo->shrinklist_lock);
3752         INIT_LIST_HEAD(&sbinfo->shrinklist);
3753
3754         sb->s_maxbytes = MAX_LFS_FILESIZE;
3755         sb->s_blocksize = PAGE_SIZE;
3756         sb->s_blocksize_bits = PAGE_SHIFT;
3757         sb->s_magic = TMPFS_MAGIC;
3758         sb->s_op = &shmem_ops;
3759         sb->s_time_gran = 1;
3760 #ifdef CONFIG_TMPFS_XATTR
3761         sb->s_xattr = shmem_xattr_handlers;
3762 #endif
3763 #ifdef CONFIG_TMPFS_POSIX_ACL
3764         sb->s_flags |= MS_POSIXACL;
3765 #endif
3766         uuid_gen(&sb->s_uuid);
3767
3768         inode = shmem_get_inode(sb, NULL, S_IFDIR | sbinfo->mode, 0, VM_NORESERVE);
3769         if (!inode)
3770                 goto failed;
3771         inode->i_uid = sbinfo->uid;
3772         inode->i_gid = sbinfo->gid;
3773         sb->s_root = d_make_root(inode);
3774         if (!sb->s_root)
3775                 goto failed;
3776         return 0;
3777
3778 failed:
3779         shmem_put_super(sb);
3780         return err;
3781 }
3782
3783 static struct kmem_cache *shmem_inode_cachep;
3784
3785 static struct inode *shmem_alloc_inode(struct super_block *sb)
3786 {
3787         struct shmem_inode_info *info;
3788         info = kmem_cache_alloc(shmem_inode_cachep, GFP_KERNEL);
3789         if (!info)
3790                 return NULL;
3791         return &info->vfs_inode;
3792 }
3793
3794 static void shmem_destroy_callback(struct rcu_head *head)
3795 {
3796         struct inode *inode = container_of(head, struct inode, i_rcu);
3797         if (S_ISLNK(inode->i_mode))
3798                 kfree(inode->i_link);
3799         kmem_cache_free(shmem_inode_cachep, SHMEM_I(inode));
3800 }
3801
3802 static void shmem_destroy_inode(struct inode *inode)
3803 {
3804         if (S_ISREG(inode->i_mode))
3805                 mpol_free_shared_policy(&SHMEM_I(inode)->policy);
3806         call_rcu(&inode->i_rcu, shmem_destroy_callback);
3807 }
3808
3809 static void shmem_init_inode(void *foo)
3810 {
3811         struct shmem_inode_info *info = foo;
3812         inode_init_once(&info->vfs_inode);
3813 }
3814
3815 static int shmem_init_inodecache(void)
3816 {
3817         shmem_inode_cachep = kmem_cache_create("shmem_inode_cache",
3818                                 sizeof(struct shmem_inode_info),
3819                                 0, SLAB_PANIC|SLAB_ACCOUNT, shmem_init_inode);
3820         return 0;
3821 }
3822
3823 static void shmem_destroy_inodecache(void)
3824 {
3825         kmem_cache_destroy(shmem_inode_cachep);
3826 }
3827
3828 static const struct address_space_operations shmem_aops = {
3829         .writepage      = shmem_writepage,
3830         .set_page_dirty = __set_page_dirty_no_writeback,
3831 #ifdef CONFIG_TMPFS
3832         .write_begin    = shmem_write_begin,
3833         .write_end      = shmem_write_end,
3834 #endif
3835 #ifdef CONFIG_MIGRATION
3836         .migratepage    = migrate_page,
3837 #endif
3838         .error_remove_page = generic_error_remove_page,
3839 };
3840
3841 static const struct file_operations shmem_file_operations = {
3842         .mmap           = shmem_mmap,
3843         .get_unmapped_area = shmem_get_unmapped_area,
3844 #ifdef CONFIG_TMPFS
3845         .llseek         = shmem_file_llseek,
3846         .read_iter      = shmem_file_read_iter,
3847         .write_iter     = generic_file_write_iter,
3848         .fsync          = noop_fsync,
3849         .splice_read    = generic_file_splice_read,
3850         .splice_write   = iter_file_splice_write,
3851         .fallocate      = shmem_fallocate,
3852 #endif
3853 };
3854
3855 static const struct inode_operations shmem_inode_operations = {
3856         .getattr        = shmem_getattr,
3857         .setattr        = shmem_setattr,
3858 #ifdef CONFIG_TMPFS_XATTR
3859         .listxattr      = shmem_listxattr,
3860         .set_acl        = simple_set_acl,
3861 #endif
3862 };
3863
3864 static const struct inode_operations shmem_dir_inode_operations = {
3865 #ifdef CONFIG_TMPFS
3866         .create         = shmem_create,
3867         .lookup         = simple_lookup,
3868         .link           = shmem_link,
3869         .unlink         = shmem_unlink,
3870         .symlink        = shmem_symlink,
3871         .mkdir          = shmem_mkdir,
3872         .rmdir          = shmem_rmdir,
3873         .mknod          = shmem_mknod,
3874         .rename         = shmem_rename2,
3875         .tmpfile        = shmem_tmpfile,
3876 #endif
3877 #ifdef CONFIG_TMPFS_XATTR
3878         .listxattr      = shmem_listxattr,
3879 #endif
3880 #ifdef CONFIG_TMPFS_POSIX_ACL
3881         .setattr        = shmem_setattr,
3882         .set_acl        = simple_set_acl,
3883 #endif
3884 };
3885
3886 static const struct inode_operations shmem_special_inode_operations = {
3887 #ifdef CONFIG_TMPFS_XATTR
3888         .listxattr      = shmem_listxattr,
3889 #endif
3890 #ifdef CONFIG_TMPFS_POSIX_ACL
3891         .setattr        = shmem_setattr,
3892         .set_acl        = simple_set_acl,
3893 #endif
3894 };
3895
3896 static const struct super_operations shmem_ops = {
3897         .alloc_inode    = shmem_alloc_inode,
3898         .destroy_inode  = shmem_destroy_inode,
3899 #ifdef CONFIG_TMPFS
3900         .statfs         = shmem_statfs,
3901         .remount_fs     = shmem_remount_fs,
3902         .show_options   = shmem_show_options,
3903 #endif
3904         .evict_inode    = shmem_evict_inode,
3905         .drop_inode     = generic_delete_inode,
3906         .put_super      = shmem_put_super,
3907 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3908         .nr_cached_objects      = shmem_unused_huge_count,
3909         .free_cached_objects    = shmem_unused_huge_scan,
3910 #endif
3911 };
3912
3913 static const struct vm_operations_struct shmem_vm_ops = {
3914         .fault          = shmem_fault,
3915         .map_pages      = filemap_map_pages,
3916 #ifdef CONFIG_NUMA
3917         .set_policy     = shmem_set_policy,
3918         .get_policy     = shmem_get_policy,
3919 #endif
3920 };
3921
3922 static struct dentry *shmem_mount(struct file_system_type *fs_type,
3923         int flags, const char *dev_name, void *data)
3924 {
3925         return mount_nodev(fs_type, flags, data, shmem_fill_super);
3926 }
3927
3928 static struct file_system_type shmem_fs_type = {
3929         .owner          = THIS_MODULE,
3930         .name           = "tmpfs",
3931         .mount          = shmem_mount,
3932         .kill_sb        = kill_litter_super,
3933         .fs_flags       = FS_USERNS_MOUNT,
3934 };
3935
3936 int __init shmem_init(void)
3937 {
3938         int error;
3939
3940         /* If rootfs called this, don't re-init */
3941         if (shmem_inode_cachep)
3942                 return 0;
3943
3944         error = shmem_init_inodecache();
3945         if (error)
3946                 goto out3;
3947
3948         error = register_filesystem(&shmem_fs_type);
3949         if (error) {
3950                 pr_err("Could not register tmpfs\n");
3951                 goto out2;
3952         }
3953
3954         shm_mnt = kern_mount(&shmem_fs_type);
3955         if (IS_ERR(shm_mnt)) {
3956                 error = PTR_ERR(shm_mnt);
3957                 pr_err("Could not kern_mount tmpfs\n");
3958                 goto out1;
3959         }
3960
3961 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3962         if (has_transparent_hugepage() && shmem_huge < SHMEM_HUGE_DENY)
3963                 SHMEM_SB(shm_mnt->mnt_sb)->huge = shmem_huge;
3964         else
3965                 shmem_huge = 0; /* just in case it was patched */
3966 #endif
3967         return 0;
3968
3969 out1:
3970         unregister_filesystem(&shmem_fs_type);
3971 out2:
3972         shmem_destroy_inodecache();
3973 out3:
3974         shm_mnt = ERR_PTR(error);
3975         return error;
3976 }
3977
3978 #if defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE) && defined(CONFIG_SYSFS)
3979 static ssize_t shmem_enabled_show(struct kobject *kobj,
3980                 struct kobj_attribute *attr, char *buf)
3981 {
3982         int values[] = {
3983                 SHMEM_HUGE_ALWAYS,
3984                 SHMEM_HUGE_WITHIN_SIZE,
3985                 SHMEM_HUGE_ADVISE,
3986                 SHMEM_HUGE_NEVER,
3987                 SHMEM_HUGE_DENY,
3988                 SHMEM_HUGE_FORCE,
3989         };
3990         int i, count;
3991
3992         for (i = 0, count = 0; i < ARRAY_SIZE(values); i++) {
3993                 const char *fmt = shmem_huge == values[i] ? "[%s] " : "%s &